A hybrid deep meta-ensemble networks with application in electric utility industry load forecasting

Abstract

Many industrial applications concern the forecasting of large quantities of related time series. This paper presents a novel hybrid modeling framework, named as Hybrid Deep Meta-Ensemble Networks (HDME-Nets), which combines local and global forecasts using meta-ensemble technology. The proposed framework can be used to generate multiple steps ahead of point and interval forecasts for a large number of time series concurrently. Our proposed framework is composed of four modules: a set of local forecasters that model each time series individually, a global forecaster that captures cross-sectional patterns with data pooling, a feature learner that extracts features from each time series in a supervised fashion, and a meta-combiner that combines the local and global forecasts according to the extracted features. The local forecasters are fitted firstly, and the other three modules are integrated seamlessly into one neural network, which is then jointly trained with a common objective measured with a custom loss function. Through testing two public available electric utility load datasets, we find that the proposed method can achieve improved forecasting performance compared against some state-of-the-art methods.