Comparison of Short-Term Load Forecasting Techniques

Abstract

This paper presents a comparative analysis of the different forecasting techniques (Statistical method (ARIMA)), Machine Learning method (Multivariate Linear Regression) and Deep Learning method (LSTM)) for short term load forecasting. The forecasting model for each of these techniques takes into account the historical load (annual), site temperature data and US calendar data as input features. The model is trained on the first nine months of data and then tested for accuracy on the remaining three months. A case study using the load profile of a commercial building (amusement park) in San Diego, California is presented, where the performance of the above forecasting techniques is compared. The error metric used for comparison is the Root Mean Squared Error (RMSE) value. The results indicate that LSTM model offers the best performance in terms of forecasting accuracy. The designed LSTM model can be deployed as part of Energy Management System (EMS) for smart grids. The robustness of the LSTM model is further explored by comparing the above LSTM encoder-decoder architecture with a standard LSTM architecture. In addition to the above dataset, both architectures were tested on two more datasets- one for an office building and another for an operations center building located in San Diego. Both architectures were trained on first 9 months of load data and tested on the remaining 3 months. The encoder-decoder architecture performed better than the standard architecture across all the datasets.