No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
Attention mechanism-aided model ensemble method of chiller energy consumption prediction
ResearchGate has not been able to resolve any citations for this publication.
Accurate savings estimations are important to promote energy efficiency projects and demonstrate their cost-effectiveness. The increasing presence of advanced metering infrastructure (AMI) in commercial buildings has resulted in a rising availability of high frequency interval data. These data can be used for a variety of energy efficiency applications such as demand response, fault detection and diagnosis, and heating, ventilation, and air conditioning (HVAC) optimization. This large amount of data has also opened the door to the use of advanced statistical learning models, which hold promise for providing accurate building baseline energy consumption predictions, and thus accurate saving estimations. The gradient boosting machine is a powerful machine learning algorithm that is gaining considerable traction in a wide range of data driven applications, such as ecology, computer vision, and biology. In the present work an energy consumption baseline modeling method based on a gradient boosting machine was proposed. To assess the performance of this method, a recently published testing procedure was used on a large dataset of 410 commercial buildings. The model training periods were varied and several prediction accuracy metrics were used to evaluate the model's performance. The results show that using the gradient boosting machine model improved the R‐squared prediction accuracy and the CV(RMSE) in more than 80 percent of the cases, when compared to an industry best practice model that is based on piecewise linear regression, and to a random forest algorithm.
The forecasting of building electricity demand is certain to play a vital role in the future power grid. Given the deployment of intermittent renewable energy sources and the ever increasing consumption of electricity, the generation of accurate building-level electricity demand forecasts will be valuable to both grid operators and building energy management systems. The literature is rich with forecasting models for individual buildings. However, an ongoing challenge is the development of a broadly applicable method for demand forecasting across geographic locations, seasons, and use-types. This paper addresses the need for a generalizable approach to electricity demand forecasting through the formulation of an ensemble learning method that performs model validation and selection in real time using a gating function. By learning from electricity demand data streams, the method requires little knowledge of energy end-use, making it well suited for real deployments. While the ensemble method is capable of incorporating complex forecasters, such as Artificial Neural Networks or Seasonal Autoregressive Integrated Moving Average models, this work will focus on employing simpler models, such as Ordinary Least Squares and k-Nearest Neighbors. By applying our method to 32 building electricity demand data sets (8 commercial and 24 residential), we generate electricity demand forecasts with a mean absolute percent error of 7.5% and 55.8% for commercial and residential buildings, respectively.
As a popular data driven method, artificial neural networks (ANNs) have been widely applied in building energy prediction field for decades. To improve the short term prediction accuracy, this paper presents a kind of optimized ANN model for hourly prediction of building electricity consumption. An improved Particle Swarm Optimization algorithm (iPSO) is applied to adjust ANN structure's weights and threshold values. The principal component analysis (PCA) is used to select the significant modeling inputs and simplify the model structure. The investigation utilizes two different historical data sets in hourly interval, which are collected from the Energy Prediction Shootout contest I and a campus building located in East China. For performance comparison, another two prediction models, ANN model and hybrid Genetic Algorithm-ANN (GA-ANN) model are also constructed in this study. The comparison results reveal that both iPSO-ANN and GA-ANN models have better accuracy than that of ANN ones. From the perspective of time consuming, the iPSO-ANN model has shorter modeling time than GA-ANN method. The proposed prediction model can be thought as an alternative technique for online prediction tasks of building electricity consumption.
General regression neural networks (GRNN) were designed and trained to investigate the feasibility of using this technology to optimize HVAC thermal energy storage in public buildings as well as office buildings. State of the art building simulation software, ESP-r, was used to generate a database covering the years 1997–2001. The software was used to calculate hourly cooling loads for three office buildings using climate records in Kuwait. The cooling load data for 1997–2000 was used for training and testing the neural networks (NN), while robustness of the trained NN was tested by applying them to a “production” data set (2001 data) that the networks have never “seen” before.Three buildings of various densities of occupancy and orientational characteristics were investigated. Parametric studies were performed to determine optimum GRNN design parameters that best predict cooling load profiles for each building. External hourly temperature readings for a 24 h period were used as network inputs, and the hourly cooling load for the next day is the output. The performance of the NN analysis was evaluated using a statistical indicator (the coefficient of multiple determination) and by statistical analysis of the error patterns, including confidence intervals of regression lines, as well as by examination of the error patterns.The results show that a properly designed NN is a powerful instrument for optimizing thermal energy storage in buildings based only on external temperature records.
Blast furnace gas (BFG) is an important energy-carrying byproduct of the iron and steel industry. High-accuracy prediction of BFG generation is the basis of the dynamic balance of gas supply–demand and energy scheduling. However, due to instrument faults, measurements of BFG are discontinuous or inaccurate, making it difficult to accurately predict future BFG generation by using historical data, which seriously restricts the development of intelligent management and coordination between various gas sources and users. To solve this problem, an attention mechanism-aided data- and knowledge-driven soft sensor is proposed to predict BFG generation. To reduce the complexity of the samples, the proposed method selects key features to simplify the model input by using attention mechanism. Genetic algorithm (GA) is used to optimize hyperparameters to improve the stability of the model. In addition, combined with the knowledge of the blast furnace process, the prediction results are reasonably constrained. The results show that the prediction accuracy of the A-DK-GA-XGBoost model is higher than that of the other prediction models, with a mean absolute error of 68.2 m³/min, a symmetric mean absolute percentage error of 0.83%, a root mean square error of 68.71 m3/min, and an R squared of 99.06%. It is proven that the A-DK-GA-XGBoost model has superior performance.
Machine learning has gained increasing popularity in building energy management due to its powerful capability and flexibility in model development as well as the rich data available in modern buildings. While machine learning is becoming more powerful, the models developed, especially artificial neural networks like Recurrent Neural Networks (RNN), are becoming more complex, resulting in “darker models” with lower model interpretability. The sophisticated inference mechanism behind machine learning prevents ordinary building professionals from understanding the models, thereby lowering trust in the predictions made. To address this, attention mechanisms have been widely implemented to improve the interpretability of deep learning; these mechanisms enable a deep learning-based model to track how different inputs influence outputs at each step of inference.
This paper proposes a novel neural network architecture with an attention mechanism for developing RNN-based building energy prediction, and investigates the effectiveness of this attention mechanism in improving the interpretability of RNN models developed for 24-hour ahead building cooling load prediction. To better understand, explain and evaluate these neural network-based building energy prediction models, the obtained attention vectors (or metric) are used to visualize the influence of different parts of model inputs on the prediction result. This helps the users to understand why predictions are made by the model, as well as how input sequences proportionally influence the output sequences. Further analysis of attention vectors can provide interesting temporal information for understanding building thermal dynamics, like the thermal inertia of the building. The proposed attention-based architecture can be implemented in developing optimal operation control strategies and improving demand and supply management. The model developed based on this architecture is assessed using real building operational data, and shows improved accuracy and interpretability over baseline models (without adopting attention mechanisms). The research results help to bridge the gap between building professionals and advanced machine learning techniques. The insights obtained can be used as guidance for the development, fine-tuning, explanation and debugging of data-driven building energy prediction models.
In recent years, the panchromatic (PAN) and the multispectral (MS) remote sensing images classification has become a research hotspot. In this paper, we propose a spatial-channel progressive fusion residual network (SCPF-ResNet) for multi-resolution remote sensing classification. Firstly, for the inputs of the proposed network, the interactive data fusion strategy (IDFS) combines generalized-intensity-hue-saturation (GIHS), and discrete wavelet transform (DWT) to interfuse patch pairs of the PAN and the MS images, so as to increase the similarity between them, thus reduce the difference in information between them. Secondly, for the branches of feature extraction, we design an adaptive spatial attention module (ASA-Module) and an adaptive channel attention module (ACA-Module) to strengthen spatial features from both larger-sized with smaller-sized targets and spectral features among channels. Finally, we insert the ASA-Module and ACA-Module into the residual modules to form a triple-branch network and use the common spatial-channel features extracted by the Fusion_Branch to gradually enhance the pure independent features extracted by the PAN_Branch and the MS_Branch, respectively. The experimental results indicate that SCPF-ResNet can achieve competitive performance.
The availability of accurate day-ahead energy prices forecasts is crucial to achieve a successful participation to liberalized electricity markets. Moreover, forecasting systems providing prediction intervals and densities (i.e. probabilistic forecasting)are fundamental to enable enhanced bidding and planning strategies considering uncertainty explicitly. Nonetheless, the vast majority of available approaches focus on point forecast. Therefore, we propose a novel methodology for probabilistic energy price forecast based on Bayesian deep learning techniques. A specific training method has been deployed to guarantee scalability to complex network architectures. Moreover, we developed a model originally supporting heteroscedasticity, thus avoiding the common homoscedastic assumption with related preprocessing effort. Experiments have been performed on two day-ahead markets characterized by different behaviors. Then, we demonstrated the capability of the proposed method to achieve robust performances in out-of-sample conditions while providing forecast uncertainty indications.
Building air-conditioning and mechanical ventilation (ACMV) systems are responsible for significant energy consumption and yet, dissatisfaction with the thermal environment is prevalent among the occupants, revealing a widespread disparity between energy-efficiency and indoor thermal-comfort in buildings. This paper presents an indoor-climate control framework that bridges this gap between energy and comfort. The framework comprises two main components: a thermal-comfort prediction model, and an optimization algorithm termed as the optimal air temperature (OAT) algorithm; they collectively act as an intelligent
mediator between the occupant and the ACMV system. Firstly, the ACMV energy consumption is modelled as a function of air temperature, and three operating frequencies of cooling components using a feedforward neural network. Secondly, the thermal-comfort prediction model predicts the thermal state index (TSI:Cool-Discomfort/Comfort/Warm-Discomfort). Thirdly, depending on the predicted TSI, the OAT algorithm locates the
optimal operating state such that Comfort state is achieved using the minimum ACMV energy consumption. Proposed framework exhibits an energy saving potential of 36.5%. It is found that 25°C is the ideal air temperature for desired comfort with minimum energy expense in the tropical buildings. Additionally, six different TSI predictive models including two general and four personal comfort models are implemented to validate the framework. The study is substantiated with extensive real human experiments in controlled thermal environment. The proposed method is scalable for its applicability with any comfort-prediction model, and adaptive for its data-driven architecture. It exhibits the potential to
achieve both occupant-comfort and energy-saving through integration with the Internet-of-Things for realizing comfort-energy balanced buildings.
Building energy consumption prediction plays an irreplaceable role in energy planning, management, and conservation. Data-driven approaches, such as artificial neural networks, support vector regression, gradient boosting regression and extreme learning machine are the most advanced methods for building energy prediction. However, owing to the high nonlinearity between inputs and outputs of building energy consumption prediction models, the aforementioned approaches require improvement with regard to the prediction accuracy, robustness, and generalization ability. To counter these shortcomings, a novel vector field-based support vector regression method is proposed in this paper. Through multi-distortions in the sample data space or high-dimensional feature space mapped by a vector field, the optimal feature space is found, in which the high nonlinearity between inputs and outputs is approximated by linearity. Hence, the proposed method ensures a high accuracy, a generalization ability, and robustness for building energy consumption prediction. A large office building in a coastal town of China is used for a case study, and its summer hourly cooling load data are used as energy consumption data. The results indicate that the proposed method achieves better performance than commonly used methods with regard to the accuracy, robustness, and generalization ability.
The increasing world population and availability of energy hungry smart devices are major reasons for alarmingly high electricity consumption in the current times. So far, various simulation tools, engineering and Artificial Intelligence based methods are being used to perform optimal electricity demand forecasting. While engineering methods use dynamic equations to forecast, the AI-based methods use historical data to predict future demand. However, modeling of nonlinear electricity demand patterns is still underdeveloped for robust solutions as the existing methods are useful only for handling short-term dependencies. Moreover, the existing methods are static in nature because they are purely historical data driven. In this paper, we propose a deep learning based framework to forecast electricity demand by taking care of long-term historical dependencies. Initially, the cluster analysis is performed on the electricity consumption data of all months to generate season based segmented data. Subsequently, load trend characterization is carried out to have a deeper insight of metadata falling into each of the clusters. Further, Long Short Term Memory network multi-input multi-output models are trained to forecast electricity demand based upon the season, day and interval data. In the present work, we have also incorporated the concept of moving window based active learning to improve prediction results. To demonstrate the applicability and effectiveness of the proposed approach, it is applied to the electricity consumption data of Union Territory Chandigarh, India. Performance of the proposed approach is evaluated by comparing the prediction results with Artificial Neural Network, Recurrent Neural Network and Support Vector Regression models.
Growing demand for energy efficient buildings requires robust models to ensure efficient performance over the evolving life cycle of the building. Energy management systems can prevent energy wastage in buildings without sacrificing occupant's comfort. However, their full capabilities have not been completely realized, partly due to their inability to quickly detect faults in HVAC systems. An accurate model and an appropriate threshold are the key factors in fault detection. The traditional method of setting a fixed threshold often leads to missed opportunities to detect faults, delayed detection of faults or false alarms. To improve the effectiveness of fault detection algorithms, we have first developed a data-driven model using extreme gradient boosting (XGBoost). We have then applied the proposed dynamic threshold method to determine occurrences of faults in real time. This method adjusts the threshold value dynamically according to the real-time moving average and moving standard deviation of the predictions. The results demonstrate the usefulness of our proposed method to detect faults early in the course. An average increase of 8.82% and 117.65%, in the F1 score, is achieved with the proposed method in comparison to the traditional fixed threshold method and an existing dynamic residual method.
The capability to forecast how differences in patterns of utilization in various kinds of loads can influence energy usage is an essential effort to decrease carbon emissions and demand-side energy management. The difference in weather change starts as the first step to change the energy consumption pattern in the domestic, commercial and industrial sector. To find the change in climate and their impact on energy usage, this study examines the medium-term (MT) and long-term (LT) energy prediction for utilities, independent power producers and industrial customers to estimate the energy usage requirement of large-scale city-wide by means of using the nonlinear autoregressive model (NARM), linear model using stepwise regression (LMSR) and random forest (least square boosting) (LSBoost) approaches, based on actual environmental as well as energy consumption data. The irregular load pattern recognition to remove the abnormal trend in actual energy usage is performed by applying the outlier detection and clustering analysis. The coefficient of variation (CV) of LSBoost model is 5.019%, 3.159%, 3.292% and 3.184% in summer, autumn, winter and spring season respectively. The machine learning (ML) techniques are validated and compared based on performance and accuracy with the previously existing Gaussian process regression (GPR) model. The optimal modeling of city-wide energy demand prediction using ML-based models are guaranteed the accurate operation and design of distributed energy systems.
Accurate building energy prediction plays an important role in improving the energy efficiency of buildings. This paper proposes a homogeneous ensemble approach, i.e., use of Random Forest (RF), for hourly building energy prediction. The approach was adopted to predict the hourly electricity usage of two educational buildings in North Central Florida. The RF models trained with different parameter settings were compared to investigate the impact of parameter setting on the prediction performance of the model. The results indicated that RF was not very sensitive to the number of variables (mtry) and using empirical mtry is preferable because it saves time and is more accurate. RF was compared with regression tree (RT) and Support Vector Regression (SVR) to validate the superiority of RF in building energy prediction. The prediction performances of RF measured by performance index (PI) were 14–25% and 5–5.5% better than RT and SVR, respectively, indicating that RF was the best prediction model in the comparison.
Moreover, an analysis based on the variable importance of RF was performed to identify the most influential features during different semesters. The results showed that the most influential features vary depending on the semester, indicating the existence of different operational conditions for the tested buildings. A further comparison between RF trained with yearly and monthly data indicated that the energy usage prediction for educational buildings could be improved by taking into consideration their energy behavior changes during different semesters.
Due to the explosion in restructuring of power markets within a deregulated economy, competitive power market needs to minimize their required generation reserve gaps. Efficient load forecasting for future demands can minimize the gap which will help in economic power generation, power operations, power construction planning and power distribution. Nowadays, neural networks are widely used for solving load forecasting problem due to its non-linear characteristics. Consequently, neural network is successfully combined with optimization techniques for finding optimal network parameters in order to reduce the forecasting error. In this paper, firstly a novel evolutionary algorithm based on follow the leader concept is developed and thereafter its performance is validated by COmparing Continuous Optimizers experimental framework on the set of 24 Black-Box Optimization Benchmarking functions with 12 state-of-art algorithms in 2-D, 3-D, 5-D, 10-D, and 20-D. The proposed algorithm outperformed all state-of-art algorithms in 20-D and ranked second in other dimensions. Further, the proposed algorithm is integrated with neural network for the proper tuning of network parameters to solve the real world problem of short term load forecasting. Through experiments on three real-world electricity load data sets namely New Pool England, New South Wales and Electric Reliability Council of Texas, we compared our proposed hybrid approach to baseline approaches and demonstrated its effectiveness in terms of predictive accuracy measures.
Energy is the lifeblood of modern societies. In the past decades, the world's energy consumption and associated CO2 emissions increased rapidly due to the increases in population and comfort demands of people. Building energy consumption prediction is essential for energy planning, management, and conservation. Data-driven models provide a practical approach to energy consumption prediction. This paper offers a review of the studies that developed data-driven building energy consumption prediction models, with a particular focus on reviewing the scopes of prediction, the data properties and the data preprocessing methods used, the machine learning algorithms utilized for prediction, and the performance measures used for evaluation. Based on this review, existing research gaps are identified and future research directions in the area of data-driven building energy consumption prediction are highlighted.
Heating, ventilation, and airconditioning (HVAC) are among the major energy demand in the buildings sector globally. Improving the energy efficiency of such systems is a critical objective for mitigating greenhouse gas emissions and transitioning towards renewable sources of energy supply. The interest of this paper is to explore means to increase the efficiency of HVAC systems in accommodating occupants' behavior in real time. For instance, rooms in office buildings are not always occupied by occupants during scheduled HVAC service periods. This offers an opportunity to reduce unnecessary energy demands of HVAC systems associated with oc-cupants' behavior. An in-depth analysis of occupants' stochastic behavior within an office building is conducted in this paper. A demand-driven control strategy is proposed that automatically responds to occupants' energy-related behavior for reducing energy consumption and maintains room temperature for occupants with similar performances as a static cooling. In this control strategy, two types of machine learning methods-unsupervised and supervised learning-are applied to learn occupants' behavior in two learning processes. The occupancy-related information learned by the algorithms is used by a set of specified rules to infer real-time room setpoints for controlling the office's space cooling system. This learning-based approach intends to reduce the need for human intervention in the cooling system's control. The proposed strategy was applied to control the cooling system of the office building under real-world conditions. Eleven case study office spaces were selected, representing three typical office uses: single person offices, multi-person offices, and meeting rooms. The experimental results report between 7% and 52% energy savings as compared to the conventionally-scheduled cooling systems.
The importance of wind power as a renewable and cost-efficient power generation technology is growing globally. The impact of wind power on the existing power system, land use, and others over time has been widely studied. Such wind integration studies, especially when they are designed as retrospective bottom-up studies, rely on detailed wind turbine data, including the geographic locations, hub height, and dates of commission. Given the frequency of gaps present in these data sets, basic concepts have been developed to cope with missing data points. In this paper, multiple advanced algorithms were compared with respect to their ability to complete such data sets. One focus was on the selection of predictor variables to analyze the impact of different completion techniques depending on the specific gaps in the data set. A sample application using a German data set indicated that random forests are particularly well suited to the problem at hand.
Back propagation neural network (BPNN) models and multiple linear regression (MLR) models are widely used to predict heating energy consumption. To improve the prediction accuracies for the BPNN and MLR models, we propose a novel sample data selection method (SDSM) combining the similar days selection with the virtual samples generation. First, a grey correlation method integrated with an entropy weight method is given to optimize the similar days selection. Then virtual samples are generated by Gaussian distribution function based on the similar samples. Finally, a new sample set (including similar and virtual samples) is obtained, and then, it is regarded as the input variable for the BPNN and MLR models. The results show that training errors and prediction errors are obviously reduced in the developed BPNN model. Although the prediction accuracy of the developed MLR model is improved by different degrees, the coefficient of determination obtained in the regression fitting process is poor. It is proved that the novel SDSM is applicable only for the BPNN model, but not for the MLR model.
A recent surge of interest in building energy consumption has generated a tremendous amount of energy data, which boosts the data-driven algorithms for broad application throughout the building industry. This article reviews the prevailing data-driven approaches used in building energy analysis under different archetypes and granularities, including those methods for prediction (artificial neural networks, support vector machines, statistical regression, decision tree and genetic algorithm) and those methods for classification (K-mean clustering, self-organizing map and hierarchy clustering). The review results demonstrate that the data-driven approaches have well addressed a large variety of building energy related applications, such as load forecasting and prediction, energy pattern profiling, regional energy-consumption mapping, benchmarking for building stocks, global retrofit strategies and guideline making etc. Significantly, this review refines a few key tasks for modification of the data-driven approaches in the context of application to building energy analysis. The conclusions drawn in this review could facilitate future micro-scale changes of energy use for a particular building through the appropriate retrofit and the inclusion of renewable energy technologies. It also paves an avenue to explore potential in macro-scale energy-reduction with consideration of customer demands. All these will be useful to establish a better long-term strategy for urban sustainability.
Data-driven models have been widely used for building cooling load prediction. However, the prediction accuracy depends not only on prediction models, but also on the selection of input variables. The aim of this study is to analyse the effect of various input variables on prediction accuracy. Eight input variables combinations are formed randomly and compared for prediction accuracy with ANN and SVM models. The training and testing data were obtained from an office building by field measurement. K-means and hierarchical clustering methods are applied to classify the input variables. Tedious information of congeneric variables is then excluded and the optimized combinations are obtained. It is concluded that the prediction models with optimized input combinations perform better than those without optimization. By comparing the different clusters of input variables, historical cooling capacity data is proved to be the most essential prediction inputs.
Building cooling load predictions can be used to better understand energy demands and to improve the energy efficiency of HVAC systems. In this study, GA-SVR and GA-WD-SVR prediction models for short-term and ultra-short-term predictions for office buildings are established. The short-term cooling load prediction model is designed to outline an HVAC system's operation strategies for the following day. The ultra-short-term cooling load prediction model is designed to inform building managers of the cooling load for the next hour and to adjust HVAC system operations in advance. An office building in Tianjin is used to train and evaluate the load prediction models. Meteorological data and one-day-ahead and one-hour-ahead cooling load records are used as model inputs. The prediction results indicate that the GA-SVR prediction model performs better for short-term cooling load prediction with MRE and R² of 6.5% and 73.1%, respectively, while the GA-WD-SVR prediction model performs better for ultra-short-term cooling load prediction with MRE and R² of 4.6% and 88.7%, respectively.
Demand Response (DR) aims at improving the operation efficiency of power plants and grids, and it constitutes an effective means of reducing grid risk during peak periods to ensure the safety of power supplies. One key challenge related to DR is the calculation of load baselines. A fair and accurate baseline serves as useful information for resource planners and system operators who wish to implement DR programs. In the meantime, baseline calculation cannot be too complex, and in most cases, only weather data input is permitted. Inspired by the strong non-linear capabilities of Support Vector Regression (SVR), this paper proposes a new SVR forecasting model with the ambient temperature of two hours before DR event as input variables. We use electricity loads for four typical office buildings as sample data to test the method. After analyzing the model prediction results, we find that the SVR model offers a higher degree of prediction accuracy and stability in short-term load forecasting compared to the other seven traditional forecasting models.
Reliable and highly-generalized prediction models of short-term electric demand are urgently needed for the building sector, as the crucial basis of sophisticated building energy management. Advances in metering technologies and machine learning methods provide both opportunities and challenges for modified approaches. With multi-resolution wavelet decomposition (MWD) as a preprocessing from the point of view of signal analysis, the hybrid support vector regression (SVR) model was applied in two case study buildings to predict the hourly electric demand intensity. Taking ten-dimensional parameters of 29 workdays as the training sample, this model was carried out in a mall and a hotel, the consumed electric demand sequences of which represented the stationary and non-stationary series respectively.
By comparisons between the hybrid SVR and the pure SVR, results indicated that the introduction of MWD can always improve the predicting accuracy for the hotel, while it is not necessary for the mall. Specifically, the similar steady level around 0.65 W/m² of absolute error was obtained for the mall and the hotel buildings, when ε was lower than 0.1. At the same time, the steady quantitative values of relative errors tended to be around 4% and 6% respectively for the hotel and the mall. Based on the limited historical readings, this paper offers an on-line prediction method of short-term electric demand, which is applicable for the further smart energy management.
Short-term building cooling load prediction is the essential foundation for many building energy management tasks, such as fault detection and diagnosis, demand-side management and control optimization. Conventional methods, which heavily rely on physical principles, have limited power in practice as their performance is subject to many physical assumptions. By contrast, data-driven methods have gained huge interests due to their flexibility in model development and the rich data available in modern buildings. The rapid development in data science has provided advanced data analytics to tackle prediction problems in a more convenient, efficient and effective way.
This paper investigates the potential of one of the most promising techniques in advanced data analytics, i.e., deep learning, in predicting 24-h ahead building cooling load profiles. Deep learning refers to a collection of machine learning algorithms which are powerful in revealing nonlinear and complex patterns in big data. Deep learning can be used either in a supervised manner to develop prediction models with given inputs and output (i.e., cooling load), or in an unsupervised manner to extract meaningful features from raw data as model inputs. This study exploits the potential of deep learning in both manners, and compares its performance in cooling load prediction with typical feature extraction methods and popular prediction techniques in the building field. The results show that deep learning can enhance the performance of building cooling load prediction, especially when used in an unsupervised manner for constructing high-level features as model inputs. Using the features extracted by unsupervised deep learning as inputs for cooling load prediction can evidently enhance the prediction performance. The findings are enlightening and could bring more flexible and effective solutions for building energy predictions.
This paper presents and discusses data-driven predictive models for the energy use of appliances. Data used include measurements of temperature and humidity sensors from a wireless network, weather from a nearby airport station and recorded energy use of lighting fixtures. The paper discusses data filtering to remove non-predictive parameters and feature ranking. Four statistical models were trained with repeated cross validation and evaluated in a testing set: (a) multiple linear regression, (b) support vector machine with radial kernel, (c) random forest and (d) gradient boosting machines (GBM). The best model (GBM) was able to explain 97% of the variance (R²) in the training set and with 57% in the testing set when using all the predictors. From the wireless network, the data from the kitchen, laundry and living room were ranked the highest in importance for the energy prediction. The prediction models with only the weather data, selected the atmospheric pressure (which is correlated to wind speed) as the most relevant weather data variable in the prediction. Therefore, atmospheric pressure may be important to include in energy prediction models and for building performance modeling.
Building energy use prediction plays an important role in building energy management and conservation as it can help us to evaluate building energy efficiency, conduct building commissioning, and detect and diagnose building system faults. Building energy prediction can be broadly classified into engineering, Artificial Intelligence (AI) based, and hybrid approaches. While engineering and hybrid approaches use thermodynamic equations to estimate energy use, the AI-based approach uses historical data to predict future energy use under constraints. Owing to the ease of use and adaptability to seek optimal solutions in a rapid manner, the AI-based approach has gained popularity in recent years. For this reason and to discuss recent developments in the AI-based approaches for building energy use prediction, this paper conducts an in-depth review of single AI-based methods such as multiple linear regression, artificial neural networks, and support vector regression, and ensemble prediction method that, by combining multiple single AI-based prediction models improves the prediction accuracy manifold. This paper elaborates the principles, applications, advantages and limitations of these AI-based prediction methods and concludes with a discussion on the future directions of the research on AI-based methods for building energy use prediction.
Forecasting heat demand is necessary for production and operation planning of district heating (DH) systems. In this study we first propose a simple regression model where the hourly outdoor temperature and wind speed forecast the heat demand. Weekly rhythm of heat consumption as a social component is added to the model to significantly improve the accuracy. The other type of model is the seasonal autoregressive integrated moving average (SARIMA) model with exogenous variables as a combination to take weather factors, and the historical heat consumption data as depending variables. One outstanding advantage of the model is that it peruses the high accuracy for both long-term and short-term forecast by considering both exogenous factors and time series. The forecasting performance of both linear regression models and time series model are evaluated based on real-life heat demand data for the city of Espoo in Finland by out-of-sample tests for the last 20 full weeks of the year. The results indicate that the proposed linear regression model (T168h) using 168-h demand pattern with midweek holidays classified as Saturdays or Sundays gives the highest accuracy and strong robustness among all the tested models based on the tested forecasting horizon and corresponding data. Considering the parsimony of the input, the ease of use and the high accuracy, the proposed T168h model is the best in practice. The heat demand forecasting model can also be developed for individual buildings if automated meter reading customer measurements are available. This would allow forecasting the heat demand based on more accurate heat consumption data directly from the customers and almost in real time. Also, the model can be used for production planning of combined heat and power (CHP) system to improve the energy efficiency.
Short-term load forecasting of building electricity usage is of great importance for anomaly detection on electricity usage pattern and management of building energy consumption in an environment where electricity pricing is dynamically determined based on the peak energy consumption. In this paper, we present a data-driven forecasting model for day-ahead electricity usage of buildings in 15-minute resolution.
By using variable importance analysis, we have selected key variables: day type indicator, time-of-day, HVAC set temperature schedule, outdoor air dry-bulb temperature, and outdoor humidity as the most important predictors for electricity consumption. This study proposes a short-term building energy usage forecasting model based on an Artificial Neural Network (ANN) model with Bayesian regularization algorithm and investigates how the network design parameters such as time delay, number of hidden neurons, and training data effect on the model capability and generality.
The results demonstrate that the proposed model with adaptive training methods is capable to predict the electricity consumption with 15-minute time intervals and the daily peak electricity usage reasonably well in a test case of a commercial building complex.
We develop a statistical machine learning framework to study the effect of eight input variables (relative compactness, surface area, wall area, roof area, overall height, orientation, glazing area, glazing area distribution) on two output variables, namely heating load (HL) and cooling load (CL), of residential buildings. We systematically investigate the association strength of each input variable with each of the output variables using a variety of classical and non-parametric statistical analysis tools, in order to identify the most strongly related input variables. Then, we compare a classical linear regression approach against a powerful state of the art nonlinear non-parametric method, random forests, to estimate HL and CL. Extensive simulations on 768 diverse residential buildings show that we can predict HL and CL with low mean absolute error deviations from the ground truth which is established using Ecotect (0.51 and 1.42, respectively). The results of this study support the feasibility of using machine learning tools to estimate building parameters as a convenient and accurate approach, as long as the requested query bears resemblance to the data actually used to train the mathematical model in the first place.
Electricity demand depends on climatic condition and the influence of weather has been widely reported in the past. The main purpose of this study is to analyse the effect of the meteorological variability on the monthly electricity demand in Italy. Temperature, wind speed, relative humidity and cloud cover are considered; the calendar effect is also taken into account. A multiple linear regression model based on calendar and weather related variables is developed to study the relationships between meteorological variables and electricity demand as well as to predict the monthly electricity demand up to 1 month ahead. The model has been extensively tested over the period 1994–2009 using different combinations of the weather related variables. Accuracies obtained are quite similar and range between 0.85% and 0.89%. Temperature turns out to be the most important variable. According to the month considered, a specific combination of the weather related variables can give the lowest Mean Absolute Percentage Error (MAPE) but differences are usually small. Good results for the summer months are obtained using Heat Index to calculate the Cooling Degree-Days; the cloud cover has a major influence from February to April.When demand forecasts are performed using the predicted meteorological variables, an overall accuracy (MAPE) around 1.3% is obtained over the period 1994–2009.The proposed model clearly identifies the influence of the weather conditions on the aggregated national electricity demand.
The energy performance in buildings is influenced by many factors, such as ambient weather conditions, building structure and characteristics, the operation of sub-level components like lighting and HVAC systems, occupancy and their behavior. This complex situation makes it very difficult to accurately implement the prediction of building energy consumption. This paper reviews recently developed models for solving this problem, which include elaborate and simplified engineering methods, statistical methods and artificial intelligence methods. Previous research work concerning these models and relevant applications are introduced. Based on the analysis of previous work, further prospects are proposed for additional research reference.
The methodology to predict building energy consumption is increasingly important for building energy baseline model development and measurement and verification protocol (MVP). This paper presents support vector machines (SVM), a new neural network algorithm, to forecast building energy consumption in the tropical region. The objective of this paper is to examine the feasibility and applicability of SVM in building load forecasting area. Four commercial buildings in Singapore are selected randomly as case studies. Weather data including monthly mean outdoor dry-bulb temperature (T0), relative humidity (RH) and global solar radiation (GSR) are taken as three input features. Mean monthly landlord utility bills are collected for developing and testing models. In addition, the performance of SVM with respect to two parameters, C and ɛ, was explored using stepwise searching method based on radial-basis function (RBF) kernel. Finally, all prediction results are found to have coefficients of variance (CV) less than 3% and percentage error (%error) within 4%.
Energy simulation was conducted for office buildings in the five major climate zones – severe cold, cold, hot summer and cold winter, mild, and hot summer and warm winter – in China using DOE-2.1E. The primary aim was to investigate the thermal and energy performance of office buildings with centralised heating, ventilation and air conditioning plants in the major climatic zones in China. The computed results were analysed in three aspects – heating load, cooling load and the corresponding building energy consumption. The building peak monthly heating load varied from 142 MW h (1033 MW h cooling) in Hong Kong to 447 MW h (832 MW h cooling) in Harbin. It was also found that passive solar designs could have large energy savings potential in the severe cold and cold climates. In Harbin, the window solar component helped lower the annual building heating load by 650 MW h. Internal loads (lighting and office equipment) and part load operations of fans and pumps also played a significant role in the overall building energy efficiency. This paper presents the work, its findings and energy efficiency implications.
In this study, the main objective is to predict buildings energy needs benefitting from orientation, insulation thickness and transparency ratio by using artificial neural networks. A backpropagation neural network has been preferred and the data have been presented to network by being normalized. The numerical applications were carried out with finite difference approach for brick walls with and without insulation of transient state one-dimensional heat conduction. Three different building samples with different form factors (FF) were selected. For each building samples 0–2.5–5–10–15 cm insulations are assumed to be applied. Orientation angles of the samples varied from 0° to 80° and the transparency ratios were chosen as 15–20–25%. A computer program written in FORTRAN was used for the calculations of energy demand and ANN toolbox of MATLAB is used for predictions. As a conclusion; when the calculated values compared with the outputs of the network, it is proven that ANN gives satisfactory results with deviation of 3.43% and successful prediction rate of 94.8–98.5%.
An essential component of a comprehensive, real-time control center for power systems is a method for the calculation of short-term load forecasts. This paper explores the use of general exponential smoothing to develop an adaptive forecasting system based on observed values of integrated hourly demand. A model is developed which offers high accuracy and operational simplicity. Forecasts of hourly MWH load with lead times of one to twenty-four hours are computed at hourly intervals throughout the week. A pragmatic viewpoint is maintained throughout the paper, which includes an analysis of two years of hourly load data, test results for the method developed, and discussions of adjustments for holidays and weather disturbances.
When summer air conditioning contributes significantly to an electric utility's system peak load, it is useful, for load forecasting purposes, to separate total system load into two components: temperature-sensitive load and nontemperature-sensitive load. Examination of historical data indicates that temperature-sensitive loads depend not only upon coincident but also antecedent weather conditions. Regression analysis techniques using a digital computer have shown that the combined effect of these weather conditions can be expressed by one composite weather variable (WV). Total system load can then be expressed as Basic Load plus the product of WV times a coefficient of air-conditioning saturation. Results obtained by application of this method to historical data (1949-1964) of Public Service Electric and Gas Company are presented.
This paper presents a new time series modeling for short term load forecasting, which can model the valuable experiences of the expert operators. This approach can accurately forecast the hourly loads of weekdays, as well as, of weekends and public holidays. It is shown that the proposed method can provide more accurate results than the conventional techniques, such as artificial neural networks or Box-Jenkins models. In addition to hourly loads, daily peak load is an important problem for dispatching centers of a power network. Most of the common load forecasting approaches do not consider this problem. It is shown that the proposed method can exactly forecast the daily peak load of a power system. Obtained results from extensive testing on Iran's power system network confirm the validity of the developed approach