A Flexible Approach for Selection of the Training Set for ANN-Based Load Forecast Using Autoencoder and Similar Day Method

This paper presents one possible approach for selecting similar days set for the artificial neural network (ANN)-based Short Term Load Forecasting method. The standard similar day forecast approach finds the best fit for the forecasted days' load based on the weather, load, or other factors. That approach is not accurate enough and finding the best fit excludes valuable information for the rest of the days in history. Moreover, if only the basic idea of this approach is used, only a narrow set of similar days is selected, which is unsuitable for ANN training. The general idea is to compare all days instead of choosing the best days. This approach provides a more flexible approach for selecting a proper training dataset for ANN. The proposed method uses an autoencoder to code all the days in history and enables comparison to the forecast day. The selection of the days is made using the Euclidian norm. The two vector distances between the forecast day code and the codes from the history are calculated using the Euclidian norm. Then the whole history is sorted by the value of the distance. Only the part with the most similar days of the initial set is used for training the ANN. Results on the test set showed that metrics improved when ANN is trained on similar days set that is selected using the proposed method. The proposed method determines enough days for the ANN training procedure and helps ANN to learn the correlation between load and other input factors optimally.