This paper examines the effectiveness of different forecasting tools for local renewable energy production, focusing on the central question: Can LSTM be used for local weather and energy forecasts and how does it compare to commercial weather forecasts?

Research question

The key to a reliable prediction of available renewable energy is a good weather forecast that includes temperature, wind speed, and solar radiation. Such forecasts that include all of these data are commercially available, such as Solcast or Meteoblue. To minimize ongoing costs, a different approach could be the usage of locally gathered data and predict the energy output with AI models. Local weather stations can improve prediction accuracy, as shown in papers by Beinert and Koutensky.

Furthermore, Wöhrle proved that it is possible to use LSTM models to provide an energy forecast up to seven days ahead on a nationwide scale. In Wöhrles’ paper the prediction scale was all Germany and the average of all german wind turbines is considered.

This paper determines whether this is also possible on a local scale and compares the results with those from commercial and satellite-based forecasts. The paper focuses on wind and solar energy as key contributors to localized sustainable energy generation that supports industrial production.

Methology

To use LSTM, local historical weather data is needed to train the model. This data was acquired in a case study by the authors with two local weather stations. As the completeness of the training data set is crucial for LSTM models, the NaN values that occur are filled with data from the nearest DWD weather station. In total, four different datasets were used:

• Self acquired data consisting weather data and the real production data from the PV-system of a local company

• Self acquired data from the roof of a building of UAS Dortmund including weather data and the output from a small wind turbine

• Weather data from DWD to complete the self acquired datasets

• Commercial weather forecast from meteoblue.com

The paper compares the predictive accuracy of LSTM models trained on historical and local data with that of conventional weather forecasts, separately, for both solar and wind scenarios.

his research is highly relevant to sustainable production, offering insights into how localized forecasting can enhance energy autonomy in industrial settings and reduce losses due to energy distribution. Preliminary results indicate that machine learning-based approaches, particularly LSTM models, hold significant promise for improving the accuracy of site-specific energy forecasts.