Comparison of Three Methods for a Weather Based Day-Ahead Load Forecasting

Day-ahead load forecasting plays an increasingly important role for the operation of networks and generation dispatch. The accuracy of the forecasting depends on many factors, including the quality and size of historical data, selected forecasting model and available information on influential factors (e.g., weather data). This paper compares three state-of-the-art models in terms of their ability to capture complex variations in load profiles and provide accurate day-ahead load forecasting: multilayer perceptron (MLP), gradient boosting regression trees (GBRT) and stacked bidirectional long short-term memory (SB-LSTM). The models are implemented on one dataset from China and another from Scotland. The performance is evaluated using mean absolute error (MAE), mean absolute percentage error (MAPE) and other energy-based indices. The presented results show that GBRT outperforms MLP with the same expert extracted load characteristics as additional inputs, but SB-LSTM provides the most accurate forecasting results, without extracting any artificial data feature from the two considered demand datasets.