Comparison of the Effect of Regularization Techniques and Lookback Window Length on Deep Learning Models in Short Term Load Forecasting

Management of electric power balance requires accurate forecasting of load and generation, especially in the context of renewable energy adoption. In this context, forecasting electric load requires more attention to decrease the uncertainties in the system operation. There have been many studies under this context, however, the effect of the lookback window for both deep learning and regularization techniques has not been fully investigated in the literature. In this study, we developed a comparative study based on 4 typical deep learning techniques, namely MLP, 1D-CNN, LSTM, and a hybrid model that is a combination of 1D-CNN and LSTM to forecast the electrical load. The effect of both regularization methods and lookback window length has been investigated in detail and found that they improved the forecasting performance based on the complexity and features of the networks. The methods are evaluated in terms of 4 different metrics namely MSE, MAE, MAPE and, R2. The results show LSTM outperformed the other methods in general, and the increase of lookback length improved its performance with the average MAPE less than 2%.