光伏发电场产量预测：基于改进元启发式优化的长短期记忆网络方法

The finite availability and unsustainable nature of fossil fuel sources have spurred growing interest in renewable energy sources. Nevertheless, substantial efforts are still required to fully incorporate energy coming from renewable sources into current power distribution networks. Although reliability plays a crucial role in enhancing the sustainability of energy production, the dependence of solar power plants on weather conditions poses challenges to maintaining consistent output without significant storage expenses. Consequently, precise forecasting of photovoltaic power generation is essential for effective grid management and energy trade market. Machine learning models have proven to be a prospective resolution due to their ability to process large datasets and capture intricate patterns within the data. This study investigates the application of metaheuristics optimization techniques to enhance light-weighted long short-term memory (LSTM) based models with and without attention for predicting power generation from photovoltaic plants. Furthermore, a modified metaheuristics optimization method based on the renowned particle swarm optimization algorithm is proposed to address the rigorous demands of hyperparameters’ optimization. Rigorous simulations on a real-world data were carried out, along with strict comparative analysis with other potent metaheuristics algorithms. A publicly available photovoltaic dataset consisting of the measurements from two plants in India was used. Additionally, this study utilized a supplementary dataset collected from a photovoltaic power plant located on the roof of Institute Mihailo Pupin (IMP) in Belgrade, Serbia. Proposed research tries to fill-in the gap in this research domain, since light-weighted LSTM models were not examined enough for this specific challenge according to the literature survey. The best produced models attained mean squared error (MSE) scores of only 0.007297 for Indian Plant 1, 0.007662 for Indian Plant 2, and 0.001812 for Institute Mihailo Pupin dataset, emphasizing considerable potential of the suggested approach for real-world applications. Finally, the applicability of the top-performance models was validated with tiny machine learning (TinyML).