ML Models Predict Wind-Solar Power Output

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Aug 6 2024

In a paper published in the journal Energies, researchers at Kyushu University developed a wind-solar tower system prototype to predict power output using various machine-learning (ML) models. Initially, linear regression was inadequate, but polynomial regression improved accuracy. After hyper-parameter tuning, deep neural networks (DNN) had the highest prediction accuracy. A 1-D convolutional neural network (CNN) also performed well, and a reduced model balanced accuracy with lower data collection requirements.

Background

Past work has highlighted the unsustainable nature of modern human consumption and the need for renewable energy sources to combat climate change. Researchers have explored various renewable energy systems, including solar chimneys and wind-solar towers, with ML and deep learning (DL) models proving effective in predicting power output. Challenges in predicting power output from wind–solar towers include handling the complex interactions between wind and solar energy sources and the need for accurate modeling despite variations in scale, weather, and data quality.

Model Assessment

This section discusses the assessment of regression models for predicting wind–solar tower (WST) output power and introduces a DL approach. The experimental data were divided into training, testing, and validation sets, with models trained and evaluated using quality metrics like the coefficient of determination (R²). Overfitting was addressed using cross-validation (CV), specifically the k-fold CV scheme, which effectively handles small datasets despite its computational cost. The section also outlines minimizing mean-square error using gradient descent as an optimization algorithm.

The overview of DL techniques highlights using artificial neural networks (ANN) and CNN for WST power prediction. ANNs consist of interconnected layers that model nonlinear input-output relationships. An ANN's activation functions, hidden layers, neurons, and other hyper-parameters are chosen to maximize performance. Fully connected networks, pooling, and convolutional layers are crucial components of CNNs, typically utilized in image processing. The layers were modified for regression tasks to handle the prediction problem.

ML Model Comparison

This section presents the results of various ML algorithms for predicting WST power output, comparing their performance based on several quality metrics and computational times. Linear regression models showed poor predictive accuracy, prompting exploring nonlinear approaches. Despite the fine-tuning efforts, ridge regression did not substantially enhance the results, which led to the decision to adopt nonlinear polynomial regression. This approach demonstrated better alignment with true values, as shown by improved prediction accuracy and lower error rates. NN and gradient boosting models were also considered, with NN showing the most promise due to their ability to model complex relationships.

Future work will involve further tuning using DL frameworks like Keras/TensorFlow for enhanced accuracy. Additionally, incorporating more features and exploring advanced optimization techniques could refine predictions. A comprehensive analysis of computational efficiency and real-world applicability will be crucial for practical implementation. Evaluating how various environmental factors and data-gathering techniques affect model performance will also be essential. Long-term research and practical testing will provide further light on the suggested models' reliability and robustness.

Model Performance

This section presents the results of DL models for predicting WST power output, focusing on feed-forward neural networks (NN) and CNN. The NN model, built using TensorFlow and Keras, demonstrated enhanced power prediction accuracy, particularly after applying regularization and dropout techniques for fine-tuning. A three-hidden-layer architecture with exponential linear unit (ELU) activation functions and an Adam optimizer yielded an R² of 0.99734.

Although CNN models featuring four convolutional layers and two fully connected layers showed impressive predictive performance, they fell short of surpassing the NN model's accuracy. Furthermore, CNN models demanded almost double the computational resources compared to the NN model. As a result, the NN model, which utilized three hidden layers, was chosen for its superior efficiency. The most accurate and computationally effective method turned out to be the NN model.

 The study also explored a reduced feature model to optimize computational resources while maintaining prediction accuracy. By selecting essential features and adding wind turbine speed as a critical variable, the reduced model achieved an R² of 0.9916, slightly lower than the full-feature NN model. This approach offers a practical solution for researchers with limited data collection capabilities or computational power while still providing reliable power predictions. The results serve as a comprehensive guide, allowing researchers to choose the most suitable model based on their specific needs, whether prioritizing accuracy, time efficiency, or resource conservation.

Conclusion

To sum up, ML and DL approaches were demonstrated to predict WST power output effectively. Linear regression showed limited accuracy, while nonlinear polynomial regression improved predictions significantly with an R² of 0.992. Among DL models, the ANN outperformed the 1-D CNN, achieving an R² of 0.997734. Future work will incorporate weather data and additional inputs to enhance prediction accuracy and optimize WST design through machine-learning algorithms.