Effect of drop-out layers inside recurrent neural networks in household load forecast application

Abstract

Ensuring precise power load forecasting is highly important in planning the secure, steady, and cost-effective functioning of the power system. Grid planning and decision-making can be based on accurate long- and short-term power load forecasting. Recently, machine learning techniques have gained widespread adoption for both long- and short-term power load forecasting. Specifically, the Long Short-Term Memory (LSTM) is customized for time series data analysis. This research proposes an LSTM model for forecasting the power load of a single house containing electrical appliances over the next 20 days. We conducted a comparative analysis of the impact of dropout layers in load forecasting applications using the LSTM model. The proposed model comprises dropout rates of 0.2, 0.3, 0.4, 0.5, and 0.6, respectively. Their impact on load forecasting is examined. The experimental results demonstrate slight variations in predictions when altering dropout layers. The results show that the effect of dropout layers on the forecast varies the accuracy by only approximately 1%. However, the models with significant dropout rates are more general than those with lower or higher rates. So the model with a dropout rate of 0.4 is suggested.