Abstract: Solar energy is characterized by clean and sustainable utilization, compared with oil and coal fossil fuels. A better environment can also be obtained to reduce the emissions of carbon dioxide (CO₂), sulfur dioxide (SO₂) and harmful substances. Three systems can be summarized in the utilization of solar energy: solar thermal, photovoltaic (PV), as well as photovoltaic combined heat and power (PV/T). Among them, the PV/T system can be expected to serve better performance and higher efficiency. In this study, a solar PV/T cogeneration system was established at Jiangxi Science and Technology Normal University (116°E, 28° 30' N) in Nanchang City, Jiangxi Province, China. The whole system consisted of photovoltaic cells, collectors, circulating water pumps and heat storage tanks. Specifically, the substrate of the photovoltaic cell was installed with the cooling copper tubes, which were closely connected using thermal grease; The pipes and interfaces that connected to the system were covered with thermal insulation materials to reduce the heat loss of the working medium during operation. In the experiment in sunny weather, the cooling water first entered the cooling copper tube to absorb the heat while dissipating heat into the photovoltaic cell, then entered the collector for the secondary heating into the heat storage tank, and finally circulated the working medium in the whole system using the circulating pump. The experimental results show that the electrical efficiency of solar PV/T system was 8.7%~12.2%, and the thermal efficiency was 51.7%. At the same time, the back propagation (BP) neural network was optimized to predict the thermal power output of the solar PV/T system, according to the improved grey Wolf optimization algorithm. A pyroelectricity model was then established to predict the solar PV/T cogeneration system using the BP neural network based on the improved grey Wolf optimization algorithm (IGWO-BP). A comparison was made between the traditional BP neural network, the BP neural network based on the particle swarm optimization (PSO-BP), and the convolutional neural network (CNN). The prediction results show that the absolute percentage error (MAPE), determination coefficient (R²), root mean square error (RMSE), efficiency factor (EF), and Pearson correlation coefficient (r) of the IGWO-BP neural network prediction model were 4.5E-05, 0.99, 0.24, 0.99 and 1.00, respectively. In the prediction of heat storage tank temperature, the indexes of the model were 8.90 E-04, 0.98, 0.07, 0.98 and 0.99, respectively. All indexes were better than the traditional BP, PSO-BP and CNN prediction model. The maximum percentage error of the IGWO-BP neural network electrical prediction model was 1.86%, which was lower than 8.78% of the BP, 5.34% of PSO-BP, and 12.41% of the CNN prediction model. The maximum percentage error of the thermal prediction model was 0.29%, which was smaller than the BP neural network of 2.45%, PSO-BP of 2.44%, and CNN of 4.28%.