Short-Term Load Forecasting using optimized LSTM Networks via Improved Bat Algorithm

Short-term load forecasting plays a preponderant role in the daily basis system's operation and planning. The state-of-the-art comprises a far-reaching set of methodologies, which are traditionally based on time -series analysis and multilayer neural networks. In particular, the existence of countless neural network architectures has highlighted its ability to cope with "hard" nonlinear approximation tasks, thus making them appropriate to perform load forecasts. Following this successful path, long short-term memory networks were employed in an optimized arrangement as forecasters, this type of recurrent neural networks has received in recent years a renewed interest for machine learning tasks. Firstly, a preprocessing stage takes place, where through the selection of similar days and correlation analysis, meaningful statistics and characteristics are extracted from the load time -series, to assemble the proper training sets. Then, Bat Algorithm is used to excel the long short-term memory network functioning, by fine-tuning its size and its learning hyperparameters. Numerical testing conducted on the Portuguese load time -series reveals promising forecasting results in an overall assessment, when compared with other state-of-the-art methods.