Seasonal electricity load forecasting using metaheuristic-optimized five-parameter sigmoid regression

Accurate annual-ahead daily electricity-load forecasting requires models that capture opposing seasonal temperature responses without sacrificing auditability. This study evaluates a seasonally partitioned five-parameter sigmoid regression calibrated with Artificial Bee Colony, Ant Colony Optimization, Firefly Algorithm, Particle Swarm Optimization, and Simulated Annealing. Models were estimated on January 2015–December 2018 data and evaluated once on the untouched 2019 period; each stochastic optimizer was run independently ten times. To establish competitive context, the same split was used for a univariate ARIMA baseline, a temperature-aware ARIMAX baseline, and a nonlinear LSTM baseline. The train-selected PSO summer–winter sigmoid combination is the operational result and achieved a 2019 MAPE of 9.790%, R 2 of 0.8004, and RMSE of 210,549, compared with MAPE values of 24.405% for ARIMA, 15.993% for ARIMAX, and 15.119% for LSTM. A test-selected AC–ABC combination reached 9.149% MAPE, but is reported only as an ex-post diagnostic and not as a deployable selection rule. Repeated-run coefficients also reveal practical non-identifiability: E and C can vary substantially, and E may reach its numerical bound without a corresponding change in predictive error. Accordingly, E is interpreted as a coupled numerical shift rather than a stand-alone physical temperature pivot. The contribution is therefore a compact, leakage-controlled seasonal forecasting and optimizer-comparison framework whose operational forecast outperforms three established baselines on this dataset.