Incremental Deep-Learning for Continuous Load Prediction in Energy Management Systems

Abstract

In this work, we introduce load prediction as continuous input for optimization models within an optimization framework for short-term control of complex energy systems. In this context, we investigated long short-term memory (LSTM) models for load prediction, because they allow incremental training in an application with continuous real-time data and have not been used in other works for continuous load prediction to our knowledge. The test and evaluation were realized using data sets of real residential data from different locations in different time resolution-hourly and minutely. Accordingly, we tested different recurrent neural network (RNN) parameters of the model such as the number of layers, the number of hidden nodes, the inclusion of regularization, and dropout in order to find the optimal LSTM configuration for our continuous load prediction application. Besides, we analyzed the quality of the LSTM algorithm by comparing it in continuous mode with the baseline model and in batch mode with the statistical model ARIMA. Training and prediction time, as well as the error stabilization time were parameters used for the evaluation. The results showed that LSTM algorithms are highly promising for integrating continuous load prediction with incremental learning.

Continual Learning for Real-World Autonomous Systems: Algorithms, Challenges and Frameworks

“…Management Systems [76] Continuous energy load prediction is highly demanding as it requires continual learning of the novel scenarios instantly. In such a case, the long short-term memory (LSTM) network predicts the energy and incrementally learns the new forecasting situations.…”