LSTM-based renewable and local energy forecasting compared to commercial forecasting tools

Accurate day-ahead forecasts of renewable energy Volatile generation are crucial for integrating variable resources into power systems and for enabling local self-consumption strategies. This study investigates Long Short-Term Memory (LSTM) neural networks as site-specific forecasting tools and benchmarks their performance against a commercial satellite-based service. Five use cases are evaluated: solar irradiance and wind speed predictions based on public data from the German Weather Service, and photovoltaic and wind power forecasts derived from on-site measurements at a small enterprise and at the University of Applied Sciences Dortmund. The models employ stacked LSTMs with adaptations such as Fourier terms to capture seasonal patterns. Results show that LSTM forecasts achieves similar overall errors as the commercial tool, particularly when trained on co-located data that reflect local siting effects. While the LSTM tends to slightly underestimate and the commercial product to overestimate, their complementary biases suggest potential in ensemble approaches. Data quality and handling of missing values remain key challenges. The findings demonstrate that LSTM-based local forecasting is a viable, cost-effective alternative to commercial services, offering enhanced accuracy and autonomy for industrial sites seeking to optimize renewable energy integration.