A Comparative Study of Multi-Scale Hybrid Deep Learning Frameworks for Estimation of Domestic Load Demand of...

In this technological era, electrical energy is the bloodstream for the economic and social development of any country. It is the need of the time that developing countries like Pakistan have strategic planning for efficient generation and utilisation of electricity and have as much cheap electricity as possible at their disposal while respecting environmental constraints. The overloaded and ageing infrastructure of an electrical power network can impact system reliability and the sustainability of power generation, transmission and distribution mechanisms. The initiation of the planning process depends upon accurate load estimation to optimally fulfil consumers' power needs. This paper compares statistical, hybrid and deep learning (DL) mechanisms, including SARIMAX, SARIMA with gradient boosting (SARIMA-GB), long short-term memory (LSTM) network, STL decomposition with LSTM, and CWT-LeNet-5-LSTM, for the prediction of residential electricity demand in the LESCO region of central Pakistan. The study uses 7670 daily feeder observations recorded between 1 January 2002 and 31 December 2022. The series is modelled at its native daily resolution and partitioned chronologically into a fitting span of 5216 days, a validation span of 920 days and a test span of 1534 days beginning 20 October 2018. All models receive the same block of 14 exogenous calendar variables, four annual Fourier harmonic pairs, day-of-week and month sine and cosine terms, a weekend indicator and a linear trend, and all neural models are trained with a validation split, early stopping and restoration of the best weights rather than for a fixed number of epochs. Accuracy is assessed with MAE, RMSE, MAPE and peak normalized RMSE under one recursive protocol at forecast leads of 1, 7, 14 and 30 days, because the ranking of the frameworks depends on the lead. Averaged over three random initialisations, the proposed CWT-LeNet-5-LSTM attains the lowest error at every multi-step lead, reaching an MAPE of 2.35 ± 0.28% at lead 30 against 3.32% for the multivariate LSTM, 3.89% for SARIMA-GB and 4.47% for SARIMAX. At lead 1, SARIMA-GB is the more accurate model (0.76% against 1.20 ± 0.17%) because the previous day's observed load dominates one-step prediction for a series whose lag-one autocorrelation is 0.978. An architecture ablation isolates the contribution of the wavelet stage, the convolutional stage, the anisotropic pooling and the calendar fusion. A stratified analysis across seasons, weekdays, weekends and high-, medium- and low-load days shows where the advantage is concentrated. Additionally, Diebold-Mariano tests identify the statistical significance of the differences.