A Machine Learning Approach to Climate‐Based Building Energy Consumption

This study proposes a data‐driven framework to assess the impact of climatic parameters on building energy consumption across three diverse Iranian cities: Tehran (semi‐arid), Bandar Abbas (hot‐humid), and Rasht (humid‐moderate). Using hourly weather data from 2000 to 2023 and Energy Plus simulations for a standardized office building, we evaluated how temperature, humidity, wind speed, and solar radiation influence annual and seasonal HVAC energy demand. A dual‐model approach was adopted, combining Long Short‐Term Memory (LSTM) and Random Forest (RF) algorithms, to forecast heating and cooling loads and quantify the relative influence of each climate variable. Results show that Bandar Abbas has the highest annual cooling load (~ 5100 kWh), while Tehran exhibits the highest heating demand (~ 5000 kWh). LSTM outperformed RF in all cities, with MAE values as low as 0.04 and RMSE reductions up to 30%. Seasonal analysis revealed that temperature and solar radiation dominate cooling loads, while wind speed and temperature influence heating demand in winter. Variable importance was derived using RF's impurity‐based feature importance and LSTM's permutation sensitivity analysis. Both methods identified temperature as the most influential factor, while wind speed had minimal impact. Model uncertainty was addressed using 5‐fold cross‐validation for RF and holdout validation with early stopping for LSTM. Practical deployment of these models is feasible via integration into IoT‐based smart HVAC systems, enabling predictive and adaptive control. The framework is transferable to other climate zones and provides a replicable approach for smart city energy planning, climate adaptation strategies, and data‐informed HVAC system design under evolving climate conditions.