Enhancing Multivariate Time Series Forecasting Accuracy Through Integration of CWT Scalograms as CNN Channels

Electricity is the essence of modern life and the main engine for it, and the presence of a sustainable electrical system has become an indispensable requirement. To ensure the continuity and quality of power grids, effective tools must be provided to manage these grids efficiently. One such tool is forecasting electrical consumption using time series. Unfortunately, modeling time series data is challenging due to its complexity and instability. Recently, deep learning has been presented as a promising technique in time series forecasting, and it has proven successful in several recent works. In this research, we present a multivariate model for short-term electrical consumption prediction by combining convolutional neural networks (CNN) with continuous wavelet transform (CWT). This approach involves transforming multivariate time series data into scalograms using CWT and feeding these scalograms into the CNN for forecasting purposes. While CNN channels are typically used with RGB components in colored images, we adapted our model for time series data by representing the channels as scalograms instead of color intensities. Each variable in the time series dataset is treated as a scalogram and inputted into one of the CNN channels for analysis. The results showed that the best suitable mother wavelet function for electrical consumption data is’Morlet’, where the trained CNN model using scalograms produced by ‘Morlet’ achieved the best performance. Furthermore, the proposed model achieved lower MAPE, MAE, and MSE compared to the baseline models and lower MAPE and MSE compared to some recently published works.