Symbolic Regression for Cooling Load Forecasting: Addressing Data Efficiency and Cold Start Challenges