Enhancing the Thermal Performance Prediction of Solar-Assisted Heat Pumps with a Two-Stage Data Generation Approach

When developing predictive models for renewable energy systems, one common problem is the insufficient quantity and quality of data. This study investigated the effectiveness of a two-stage generative adversarial network (GAN) model as a potential solution. It employs a limited set of experimental data to predict the coefficient of performance (COP) of a solar-assisted PVT (photovoltaic-thermal) heat pump. The model uses a baseline and a fine-tuned CTGAN to generate input variables and then applies a deep learning-based regression model to predict the output. Finally, a dataset combining both real and generated data was created, and a new regression model was trained on it. The importance of the two-stage approach and the need for fine-tuning the model to enhance modeling performance is demonstrated in evaluation of the GAN, including the Kolmogorov-Smirnov test, Wasserstein distance, t-SNE (t-distributed stochastic neighbor embedding) diagram, and regression metrics. After applying Bayesian optimization to enhance structural performance, the RMSE was reduced from 1.061 to 0.889. In particular, these results were lower than the RMSE of 1.086 obtained using the traditional one-stage CTGAN. Although the predicted output variables indicate a favorable level of precision, they closely approach, but do not fully match, the accuracy of a traditional simulation method. However, considering the minimal amount of experimental data and the benefits of the model in terms of reduced time and cost, they can be considered satisfactory, especially when compared with costly simulation methods that require complex mathematical modeling.