Optimum input sequence size for a sliding window-based LSTM neural network used in short-term electrical load forecasting

Proper input sequence size selection has a major impact on the accuracy of short-term electrical load forecasting when using recurrent deep Long Short-Term Memory (LSTM) neural networks. A sliding window approach is frequently used for segmenting input data in models of this type. This paper aims to investigate several input sequence sizes in the given context assuming that the best result will coincide with a high value of the autocorrelation function. Therefore, a certain number of input sequence sizes was first determined using autocorrelation analysis. 5 datasets were formed based on historical data on electricity consumption for 5 randomly selected European countries. Observations were monitored on an hourly basis. For each dataset, after model training, an evaluation of the test data set was performed. Patterns identified by the autocorrelation function were recognized by the recurrent LSTM. It was shown that the most accurate result coincides with a point that has a high value of the autocorrelation function. More specifically, the best result was obtained for the size of the input sequence of 168 hours.