In a paper published in the journal Energies, researchers estimated the energy production of wind and solar power plants using deep learning (DL) methods. For solar plants, long short-term memory (LSTM) achieved higher accuracy than seasonal autoregressive integrated moving average (SARIMA).
Study: Deep Learning Boosts Renewable Energy Forecasting. Image Credit: metamorworks/Shutterstock.com
For wind energy, the convolutional neural networking—gated recurrent unit (CNN-GRU) achieved the highest accuracy in forecasting wind speeds, surpassing other methods.
Related Work
Past work has explored various methods for forecasting renewable energy production using artificial intelligence (AI) and DL, including solar and wind energy. Despite advancements, challenges such as high initial costs, variability in energy sources, and the need for accurate, long-term predictions persist. These issues complicate efficiency and cost-effectiveness in renewable energy systems.
Findings and Analysis
The study utilized LSTM and SARIMA models to analyze univariate time series data for renewable energy forecasting. Performance was assessed using metrics such as mean square error (MSE), root mean square error (RMSE), normalized MSE (NMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE).
For the LSTM model, the sequential basis from the Keras library was employed, featuring four layers with specific configurations: 32 inputs in the first layer, a tracking layer with a value of 0.5, an intermediate LSTM layer with 25 entries, and a dense output layer with one unit. The rectified linear unit (ReLU) activation function and RMSProp algorithm were used, with 16 and 200 epochs batch size.
The SARIMA model was developed using Python's Statsmodels library and the pmdarima library for optimal parameter selection. The team used the akaike information criterion (AIC) method to choose the best model. The Dickey-Fuller unit root test was also employed to ensure stationarity in the time series data.
The study examined three datasets from solar power plants, each with 1 MW capacity and spanning three years of daily electricity production data. Wind speed data from two locations were also analyzed, with two years of data from meteorological stations. The study carefully selected the solar and wind energy system locations for analysis.
Future Research Directions
The study analyzed three solar energy systems, each with a 1 MW capacity, by making one-year forecasts for each. The data were split into 70% for training and 30% for testing and applied to both the SARIMA and LSTM models. Performance evaluation used MSE, RMSE, NMSE, MAE, and MAPE metrics. Analysis revealed that the LSTM model consistently outperformed the SARIMA model. For instance, LSTM predictions closely matched actual energy production for Plant_A, with NMSE of 0.05 and MAPE of 81%, compared to SARIMA's NMSE and MAPE results.
Similarly, the LSTM model outperformed SARIMA in predicting energy production for plant_B and plant_C. With accuracy rates of 65% for plant_B and 59% for plant_C, respectively, the LSTM model produced predictions very close to the actual data, while SARIMA's accuracy rates were 49% and 41%, respectively. Many sudden rises marked the discrepancies in SARIMA's predictions and falls, making the LSTM model's smoother results more reliable.
A K-fold cross-validation with k=5 was performed to validate the models further, dividing the three-year dataset into five parts. The LSTM model consistently showed superior performance across all folds, achieving high accuracy compared to SARIMA. Annual energy production estimates for the three solar power plants confirmed that LSTM achieved an accuracy rate of 81% in the best case. At the same time, SARIMA reached a maximum of 66%, with the lowest accuracy for SARIMA at 41%.
For wind energy forecasting, various models were evaluated for weekly, monthly, and annual forecasts, including CNN-GRU, GRU, LSTM, LSTM-GRU, CNN-LSTM, and CNN-RNN. The CNN-GRU hybrid model achieved the lowest RMSE values, with 0.0301 for weekly and 0.0127 for monthly forecasts. The proposed hybrid model for annual forecasts delivered an R2 value of 0.9941, outperforming other models, with CNN-LSTM following closely behind.
In examining a 12-week dataset, all DL models showed close performance, though variability was present. The CNN-GRU model performed better across the evaluation period, providing more accurate forecasts on average than other models. It highlights the robustness and effectiveness of the CNN-GRU model in predicting wind energy over short-term periods.
Conclusion
To sum up, this study focused on estimating electricity production from solar power plants using time-series analysis with DL methods LSTM and SARIMA. The analysis was conducted on datasets from three different solar power plants, with performance measured using RMSE, MSE, NMSE, MAE, and MAPE. The results showed that the LSTM model outperformed SARIMA, benefiting from GPU support and facilitating faster and more consistent results. The SARIMA model, reliant on the CPU, achieved a different level of performance.
Wind speed data, sourced from a meteorological station over three years, were used to evaluate predictions across various models, including CNN-LSTM, CNN-RNN, LSTM-GRU, GRU, LSTM, and CNN-GRU. The LSTM-GRU model demonstrated superior accuracy and reliability in forecasting weekly, monthly, and annual wind speeds with high-performance metrics (MAPE: 0.5348, RMSE: 0.1095, R2: 0.9981). Future work aims to enhance the model for monthly forecasts, enabling quicker problem detection and more efficient maintenance scheduling, ultimately reducing time and costs in energy production.
Deep Learning Secures IoT with Federated Learning