A Generic Load Forecasting Method for Aggregated Thermostatically Controlled Loads Based on Convolutional Neural Networks

Abstract

Thermostatically Controlled Loads (TCLs) are excellent candidates for demand response. Load forecasting of aggregated TCLs is used to help utility companies predict and manage their load demand. This paper presents a generic load forecasting method that aims at providing accurate forecast for different TCL datasets without the need of extracting specific predictors. It relies on Data Quantization and Dequantization modules to processes the inputs and outputs, and a group of Convolutional Neural Network modules used for automatic feature learning and multi-horizon load demand forecasting. The forecasting method was studied with different simulated TCLs data and aggregation sizes resulting in an enhanced performance when compared with three traditional forecasting methods. Additionally, the generalization capacity of the proposed forecasting method was studied with real data from a conference environment (building) corroborating the performance advantage of the method and the generalization capacity of the "predictor-less" approach.