Random Hyper-parameter Search-Based Deep Neural Network for Power Consumption Forecasting

In this paper, we introduce a deep learning approach, based on feed-forward neural networks, for big data time series forecasting with arbitrary prediction horizons. We firstly propose a random search to tune the multiple hyper-parameters involved in the method performance. There is a twofold objective for this search: firstly, to improve the forecasts and, secondly, to decrease the learning time. Next, we propose a procedure based on moving averages to smooth the predictions obtained by the different models considered for each value of the prediction horizon. We conduct a comprehensive evaluation using a real-world dataset composed of electricity consumption in Spain, evaluating accuracy and comparing the performance of the proposed deep learning with a grid search and a random search without applying smoothing. Reported results show that a random search produces competitive accuracy results generating a smaller number of models, and the smoothing process reduces the forecasting error.