•SL, DL and DRL models were constructed for multi performance prediction of gymnasiums under different weather conditions.•Accuracy, time, interpretability, ease-of-use, hyperparameters, and robustness of the metamodels have been investigated.•Energy consumption, temperature and CO2 time series vectors prediction were modeling with nine training sets.•Methodology for hyperparameters comparison using exhaustive grid search and sensitivity analysis.•RDPG is most accurate but time-consuming. LSTM is more efficient and also has relatively excellent accuracy. Building performance simulation (BPS) is essential for testing energy demand and indoor environment quality of different building designs. However, software for BPS is computationally intensive and impractical to run thousands even millions of simulations for performance analysis and optimization. Especially for the large space buildings, which usually have complex forms and high energy consumption. The computational problem could be overcome by the adoption of metamodels. Most correlational research focuses on single performance prediction of a particular building, which makes the model less robust when applied to multi vectors prediction for different buildings under multiple weather conditions. This paper leveraged six metamodels, which include recurrent deterministic policy gradient (RDPG), asynchronous advantage actor-critic (A3C), long short-term memory (LSTM), convolution neural network (CNN), artificial neural network (ANN) and support vector regression (SVR), to predict hourly-based multi performance vectors of gymnasiums under various design parameters and multiple weather conditions. Six metamodels are trained and tested on a large scale of datasets simulated by EnergyPlus over four gymnasium cases in different cities of China. The accuracy, efficiency, ease-of-use, robustness and interpretability of the models are investigated. To conduct a fair and detailed comparison, a methodological approach using grid searches for model settings selection assisted by sensitivity analysis is pursued. Principal component analysis (PCA) is also adopted to interpret the work process of the metamodels. The comparison showed that the RDPG model provides the most accurate prediction results with R2 converges at 0.993, 0.982 and 0.941 for energy, temperature and CO2, respectively. LSTM model is more efficient than RDPG, and suitable for users who need emphasis on both time and accuracy. ANN is suitable for users with limited time and require models of ease-of-use and robustness. SVR and ANN could be used for the automatically co-simulation with BPS software. In future research, the influence of occupant behavior also should be investigated.