New input identification and artificial intelligence based techniques for load prediction in commercial building

Abstract

The accuracy of prediction models for electrical loads are important as the predicted result can affect processes related to energy management such as maintenance planning, decision-making processes, as well as cost and energy savings. The studies on improving load prediction accuracy using Least Squares Support Vector Machine (LSSVM) are widely carried out by optimizing the LSSVM hyper-parameter which includes the Kernel parameter and the regularization parameter. However, studies on the effects of input data determination for the LSSVM have not widely tested by researchers. This research developed an input selection technique using Modified Group Method of Data Handling (MGMDH) to improve the accuracy of buildings load forecasting. In addition, a new cascaded Group Method of Data Handing (GMDH) and LSSVM (GMDH-LSSVM) model is developed for electrical load prediction to improve the prediction accuracy of LSSVM model. To further improve the prediction model ability, a Modified GMDH has been cascaded to the LSSVM model to enhance the accuracy of building electrical load prediction and reduce the complexity of GMDH model. The proposed models are compared with GMDH model, LSSVM model and Artificial Neural Network (ANN) model to observe the prediction performance. The performances of prediction for each tested models are evaluated using the Mean Absolute Percentage Error (MAPE). In this analysis, the proposed prediction model, gives 33.82% improvement of prediction accuracy as compared to LSSVM model. From this research, it can be concluded that cascading the models can improve the prediction accuracy and the proposed models can be used to predict building electrical loads.