Quality analysis of combined similar day and day ahead short-term load forecasting using recurrent neural networks

A novel approach for forecasting short-term loads using recurrent neural networks (RNNs) is proposed to reduce the forecast error. Conventionally, RNN will have a single network framework trained either on a similar day (SD) or on the day ahead (DA) approach. A model trained either on a similar day or the day ahead can only learn the characteristics of either method. Hence in this paper, an optimal architecture of a recurrent network model is developed that utilizes both similar day and day ahead features for effective load forecasting. Statistical analysis is carried out on different network models such as long short-term memory and gated recurrent unit based networks, for optimal selection of layers and nodes. Based on the statistical analysis, an optimal network model is developed using both day ahead and similar day features and is compared against the day ahead approach and similar day approach. The effectiveness of the proposed network models are tested using Global Energy Forecasting Competition load data.