Comparative analysis of long-term electricity load forecasting using machine learning and meta-heuristic-optimized hybrid architectures

Accurate long-term electricity peak load forecasting is essential for modern power system growth, reliability, and effective infrastructure planning. There is no single model for load forecasting that works for all electric utilities, as it can differ from region to region and among customer types. This paper presents a strong, data-driven approach that uses historical economic and demographic data from 2016 to 2025 to predict electrical power loads over a 10-year period for Adigrat city. Using advanced Pandas techniques for imputation of missing data and customer class correlation analysis to identify strongly related attributes, data pretreatment was meticulously carried out. In order to tackle the forecasting problem, we compare our proposed novel hybrid framework Particle Swarm Optimized (PSO) LSTM-NN with Machine Learning (ML) regression, Long Short-Term Memory (LSTM), LSTM-Convolutional Neural Network (CNN), and a hybrid LSTM-(NN)-Gated Recurrent Unit (GRU) model. To address potential out-of-sample scaling issues and risks of performance drop with standard standalone deep learners, we introduce an intelligent input-combination conditional feature selection engine. The uniqueness of this research comes from creating a highly robust, customer-focused framework for long-term load forecasting (LTLF) which combines tailored demographic analysis, automated handling of missing data, and an improved hybrid deep learning structure. This work primarily contributes a stringent preprocessing pipeline (featuring missing data mitigation and multivariate correlation analysis) alongside empirical proof that the LSTM-NN-PSO model delivers superior accuracy and extended stability when projecting intricate macroeconomic variables across a 10-year period.