Machine learning-based solar power forecasting device and method that does not use future weather forecast data

Solar Power Generation Forecasting Using Pre-Peak Measured Weather Data and Neural Network Ensembles

This technology is a machine learning-based device and method that predicts solar power generation during peak hours by using temperature, humidity, cloud cover, and solar radiation data measured before the peak time, rather than relying on future weather forecasts. It trains three neural network models—RNN, LSTM, and GRU—and calculates the final power output using a weighted average based on their prediction accuracy.

Predicting solar power generation during peak hours is critical for pricing, grid stability, and power planning. However, existing methods rely on future weather forecasts for peak periods. Because weather forecasts are inherently uncertain and prone to errors, predictions based on them often struggle to achieve high accuracy.

This technology uses only weather data already measured earlier in the day, eliminating the impact of forecast errors. By using an ensemble that weights different recurrent neural network models based on their accuracy, it provides more stable predictions than a single model. The results are displayed as unit-time power generation, allowing for immediate use by managers. This technology can be applied to solar plant operations, power trading bids, and energy storage system scheduling, enabling reliable peak-hour generation estimates without external forecast services.

‍

‍

Key Features:

• A measurement table maintenance unit that keeps a data table recording daily sets of weather data measured at unit intervals before peak hours and solar power generation data during peak hours
• A configuration that randomly classifies daily data so that the training set is larger than the test set and generates three prediction models based on RNN, LSTM, and GRU through supervised learning
• A weight assignment unit that measures the accuracy of each prediction model using the test set and assigns weights that sum to 1 based on the accuracy ratio
• A prediction unit that inputs the day's pre-peak temperature, humidity, cloud cover, and solar radiation data into the three models and calculates the weighted average of the outputs as the peak-hour power generation

‍

This invention was developed with support from the Ministry of Education's research on deep learning-based intelligent space decision-making models for the global optimization of human-space interaction.

‍