This paper conducts a systematic comparison of several hybrid models for short‑term building thermal load forecasting. First, it surveys common techniques reported in the literature and finds that meta‑heuristic algorithms or data‑driven models are typically employed to calibrate the parameters of a base model. Then each method’s input‑output features, main strengths, and drawbacks are qualitatively evaluated. Using a historical thermal demand dataset from Scottish households together with corresponding weather forecasts, the performance of 13 existing hybrid approaches is empirically tested. The Empirical Mode Decomposition‑Long Short‑Term Memory‑Markov (EMD‑LSTM‑Markov) model achieves the highest day‑ahead accuracy for heating and domestic hot water (DHW) power patterns, capturing local peaks well but underestimating large power surges. In contrast, methods such as Support Vector Machine‑Simulated Annealing (SVM‑SA) and Random Forest‑Improved Sparrow Search Algorithm‑LSTM (RF‑ISSA‑LSTM) produce smoother demand profiles and tend to underestimate rapid peak changes.
Review: By integrating decomposition, deep learning, and probabilistic transition, hybrid models markedly improve the granularity of short‑term thermal load forecasts, yet dedicated error‑compensation strategies are still needed to handle extreme power spikes.