Electricity load forecasting using advanced feature selection and optimal deep learning model for the variable refrigerant flow systems

Abstract

The control optimization of a variable refrigerant flow (VRF) system requires an accurate electricity load forecast because VRF systems have a wide range of energy consumption owing to part load ratios. Currently, the empirical gray and black-box models are widely used for electricity load forecasting and may not capture the non-linear and dynamic characteristics of VRF system. This paper presents a long short-term memory based sequence-to-sequence (seq2seq) model to forecast the multi-step ahead electric consumption of VRF systems according to the state information and control signals. Increasing the depth of the network and the number of neurons per hidden layer cannot improve the performance of the proposed model for testing data up to a limited number of layers, indicating overfitting. This paper presents two methods to address this limitation. First, the feature selection methods were implemented resulting in computationally efficient models with higher accuracies. Pearson correlation and random forest methods were used to identify the relationship among features and thus ascertain both relevant and redundant features. In the second approach, a Bayesian optimization is presented to identify the hyperparameters of a given model that improve the performance of load forecasting. The results demonstrate that the optimized seq2seq model with feature selection is capable to predict the electricity consumption and the daily peak electricity usage reasonably well in a test case of a commercial building with VRF systems. The accurate and robust load forecasting model enables building operators to simulate the operating configurations of VRF system without making physical changes.