“The SFPV and GSPV systems are installed and tested under the same climatic conditions of Azizia, Kingdom of Saudi Arabia and in-depth evaluated with respect to produced electric power, PV surface panel temperature, PV-DC-voltage, PV-DC current, and energy yield and efficiency,” the team explained. “The second objective of this study focuses on the application of advanced artificial intelligence models for forecasting electrical power generation and PV surface panel temperature in both SFPV and GSPV systems, an area that has not been rarely investigated.”
Both the SFPV and the GSPV setups consisted of two bifacial panels with a max power of 545 W. Both setups also included an inverter, a battery, and a set of data loggers and measurement devices. The SFPV system was installed 25 m from the Bahrain Gulf Coast in Azizia at a depth of 1.5 m, while the GSPV was installed nearby on land. The SFPV also used a wooden frame, reusable plastic drums, a support structure made of stainless steel, steel ropes, hooks, and anchoring concrete blocks.
The analysis showed that the average ambient temperature over the long term fluctuated from 15.35 C in January to 36.0 C in July, with relative humidity reaching 31.65% in June and peaking at 68.23% in December. The global horizontal solar intensity on a daily basis ranged from 3.30 kWh/ m2/day to 7.74 kWh/m2/day, with an overall average of 5.64 kWh/m2/day throughout the year. Furthermore, the average wind speed at 10 m above sea level varied between 3.71 m/s in October and 5.42 m/s in June.
Measurements of both devices were made in June 2024 and showed the SFPV system improved the average PV electrical power and accumulated net daily electrical energy by 59.25% and 69.70%, respectively, compared to the ground-mounted system. That was due, in part, to the cooling effect of the seawater. While the average measured on the surface of the GSPV was 58.40 C, the SFPV had a 39.5 C, a reduction of 32.36%.
To predict the capabilities of these systems, the group has combined the brown-bear optimization algorithm (BBOA) with the long short-term memory (LSTM) technique. BBOA is inspired by the natural behaviors of brown bears and is used to fine-tune the hyperparameters of the LSTM model. Hyperparameters are the external configurations set before the LSTM learning process begins, governing its operation. The LTSM then uses its ability to understand patterns to forecast the results.
“Dataset partitioning was done using 70/30 splits, in which 70% of the dataset was allocated for training, and 30% for testing,” the group explained. “The input variables for the model include features such as time, solar radiation, PV current, PV voltage and ambient temperature, while the target outputs are electrical power and PV surface temperature.”
The LSTM-BBOA was then benchmarked against three other models: light gradient-boosting machine (LightGBM), LSTM alone, and gated recurrent unit (GRU). According to the results, the LSTM-BBOA model achieved superior robustness in both SFPV and GSPV systems. In the case of the SFPV's electricity, it achieved a deterministic coefficient (R²) of 0.9998. For compression, the LSTM alone received 0.9966, and the LightGBM had 0.9844.
The analysis showed that the hybrid LSTM-BBOA revealed a “robust” performance with minimal mean absolute error (MAE), root mean square error (RMSE), and coefficient of variation (COV) values of 0.4884, 0.5031, and 0.1938 for the SFPV power production predictions. While, the standalone LightGBM exhibited the maximal MAE, RMSE, and COV values of 5.7036, 12.6872, and 20.3577, respectively.
“The LSTM-BBOA model achieved peak efficiency coefficient (EC) and overall index (OI) values of 0.9998 and 0.9931, respectively, exceeding the LSTM model's scores of 0.9969 and 0.9472 for SFPV power production,” the scientists concluded. “In comparison, LightGBM recorded the lowest EC and OI values, at 0.9844 and 0.9190, respectively.”
