Utility companies strategy for short-term energy demand forecasting using machine learning based models

This paper presents machine learning based models applied for forecasting the future energy requirement of water source heat pumps. From the model analysis, four machine learning based models were accrued which are: i) CDT; ii) FitcKnn; iii) LRM; and iv) Stepwise-LRM. The input parameters include the environment data, power usage data of the water source heat pump, hour-type/day-type, and the output, (the output being the net electricity usage of the water source heat pump). The forecasting session is divided into two parts: i) weekly; and ii) monthly to measure the model's performance for short and medium-term perceptive. The performance evaluation statistics expandability for assessing the model's performance are the MAE, RMSE, and MAPE. As a result, the MAPE and MAE for the subsequent month energy forecasting of the ML models; (CDT, FitcKnn, LRM, and Stepwise-LRM) are 0.044%, 0.051%, 0.776%, 0.343% and 7.523, 1.766, 12.317, 5.969 respectively. To verify the accuracy of the forecasts given by the proposed models, four existed validation methods - the BRNN, LMA, TB and GPR are applied and the forecasting performance and efficiency contrasted with the proposed machine learning models. The outcomes from this analysis can accommodate to recognize the significance of climatic variables on power usage within a building background and enhance the forecasting efficiency of short and medium-term energy forecasting with inadequate climatic data, power efficiency retrofit, and facilities for investment by utilities, industrial and commercial consumers.