Wavelet-based Decomposition and Artificial Neural Networks for Short-Term Prediction of Power Demand in District Heating

International audienceAs part of the second phase of the OptiEnR research project, the present work deals with improving the operation of multi-energy district boilers, by adding optimally designed and controlled thermal storage tanks to the plants. Multi-energy district boilers are connected to local heat networks for thermal energy distribution. The aim of the project is so to develop a Model-based Predictive Controller (MPC) in order to optimize, over a prediction horizon, the use of the tank. The controller will produce optimal command sequences in order to regulate the amount of thermal energy to be stored or released. The development of such a controller requires the power demand to be forecasted accurately. As a result, we propose in this paper a short-term forecast methodology built using a Wavelet-based Multi-Resolution Analysis (WMRA) and multilayer Artificial Neural Networks (ANN). The main idea behind such a methodology is to replace the prediction of the original time series, which is characterized by a high variability, by the prediction of its wavelet coefficients (of lower variability). The WMRA allows a time series to be decomposed into approximations (i.e. low-frequency coefficients) and details (i.e. high-frequency coefficients) using a filter bank, while preserving its main temporal characteristics. So, we decomposed observation sequences (of 24 hours) and used self-growing artificial neural networks trained with the cascade-correlation algorithm in order to forecast the coefficients. The future values of power demand are then obtained by summing these coefficients. As a key point, we carried out a complete parametric analysis in order to find the optimal configuration with respect to the wavelet type and order, decomposition level, right number of observation sequences and topology of the artificial neural networks. The results we obtained highlight the ability of the proposed methodology to forecast power demand with high accuracy