Scientists from the International Institute for Water and Environmental Engineering at Thomas Sankara University and Joseph Ki-Zerbo University in Burkina Faso have developed a system for forecasting solar energy production at the 30-megawatt Nagréongo power plant. Using artificial intelligence and machine learning technologies, the researchers achieved high precision in predicting energy production fluctuations depending on weather conditions.
Like other countries in the Sahel region, Burkina Faso considers solar energy as the main pathway to energy independence. However, production at solar power plants is known to be highly dependent on cloud cover, dust and seasonal variations in solar radiation. Any sudden reduction in solar flux leads to voltage surges in the grid, complicating balancing processes. Aiming to solve this problem, the scientists turned to machine learning systems.
The researchers trained several neural networks: a recurrent network, a long short-term memory network, a network with controlled recurrent units and a hybrid network. They analyzed data from the Nagréongo power plant, which were collected every five minutes throughout 2024. This way, the algorithms learned to predict power generation under different climatic conditions, from the dry and hot season to the rainy season. Prior to training, the data were cleaned up and normalized according to international standards to eliminate noise, omissions and measurement errors.
In the dry season, the model with controlled recurrent units produced the most accurate forecasts, with an error of approximately 4%. During hot months, the network with long short-term memory proved the best, with an error of no more than 2%. In the rainy season, the hybrid model proved the most robust, maintaining accuracy even with significant fluctuations in illumination. Thus, the scientists demonstrated that the models can be selected based on the specific season, which is especially important in regions with sharp climate contrasts.
The research team also assessed the technical efficiency of the power plant. Its efficiency factor rose from 73.7% in 2023 to 77.4% in 2024, reflecting increased solar panel efficiency. The improvement was driven by the implementation of a system for regular dust cleaning, the lifting of power supply restrictions after grid line upgrades and reduced equipment downtime.
The neural networks have given power plant operators the opportunity to predict changes in power generation in near real time. An error of 2–5% of the nominal power is very low for solar plants operating in a variable cloud cover and dusty Sahel winds. This makes it possible to adjust the load in advance and provide grid stability without interruptions or energy losses.
In the future, the researchers plan to supplement the model with data from satellites and weather stations to forecast power generation not only for the next few minutes but also for the next 24 hours.
