CNN-N-BEATS: Novel Hybrid Model for Time-Series Forecasting

Time-series forecasting (TS) is a vital tool for scientific study and has many applications in a wide range of disciplines, including engineering, economics, finance, environmental science, weather, energy, etc. Unquestionably, the ability to accurately forecast trends and patterns from historical load data in particular, is a benefit, that is especially helpful in decision-making scenarios and aids in understanding the underlying mechanisms connected to observable events. This paper aims to tackle this problem by proposing a novel hybrid forecasting model for time-series forecasting. The proposed forecasting model consists of a Convolution Neural Network (CNN) coupled with a Neural Basis Expansion Analysis for interpretable Time-Series forecasting model (N-BEATS) dubbed CNN-N-BEATS. The CNN-N-BEATS model outperforms popular forecasting techniques as it proposes a mechanism for dealing with a plaguing problem associated with time-series forecasting, that is, the distribution shift that tend to occur during forecasting. In addition, we also implement a technique, called Weighted Linear Stacking (WLS), for improving time-series accuracy when a large amount of historical load is available. The WLS technique combines the output of numerous forecasting models in order to achieve better accuracy. This research also shows that load forecasting using only historical data can produce results that are as accurate as those achieved by leveraging additional exogenous variables during the forecasting process.