Improving solar power forecasting with an ensemble of deep learning models

Machine Learning


Artificial intelligence helps predict power generation from volatile renewable energy sources like wind and solar. Machine learning and (a subset of) deep learning are beginning to replace traditional weather and satellite data-based forecasting and statistical prediction models.

Deep learning models, ranging from simple artificial neural networks to complex “long short-term memories” (LSTMs, an architecture that is particularly effective at making predictions based on continuous data), have been introduced to improve the accuracy of predictions.

Now, three researchers from the Indian Institute of Technology, Kolkata — Rakesh Mondal, Surajit Kumar Roy, and Chandan Giri — have come up with an improved AI technique for forecasting solar power. Instead of using a simple deep learning model, these scientists employed an ensemble of deep learning models that “go a step further than simple deep learning models,” which they say has resulted in improved accuracy.

The benefits of AI

Ensemble models, which combine predictions from multiple individual deep learning models, aren't entirely new. In their Energy paper, the authors acknowledge that other researchers have tried ensemble modeling approaches, but say their work “incorporated features that improve the accuracy of predictions.” These features include parameters such as the physical characteristics of a solar panel, such as the number of cells in the panel, the panel's maximum operating temperature, material type, and ambient temperature. “No existing methodology takes these parameters into account when forecasting solar power production,” the authors say.

Mondal, Roy and Giri used a technique called “bidirectional long short-term memory,” or BI-LSTM, which is a type of recurrent neural network (RNN) designed to process continuous data. Unlike standard LSTMs, which process data in one direction (from past to future), BI-LSTMs process data in both directions (from past to future and future to past). This allows the model to take into account both past and future information to better understand the context.

The researchers prepared the dataset by combining meteorological parameters and solar power generation data, and enriched the dataset with meteorological data and the physical characteristics of the solar panels installed at each solar power plant. According to the researchers, the BI-LSTM model is able to predict the future solar power generation of a particular solar power plant in both the short and long term, regardless of the geographic location of the solar power plant.

“Short-term forecasts allow us to predict solar power generation 15 minutes to one hour in advance, while long-term forecasts allow us to predict solar power generation 1-3 days in advance with remarkable accuracy,” the paper states.

Mondal, Roy and Giri compared the results of their proposed model with existing datasets and multiple standard deep learning models and found that “our model performed better than the conventional models.” They also validated their model using various solar power plants in Durgapur, India. “Our model also outperformed the base models in long-term predictions.”

From data to decisions

In response to email Quantum Dr Giri said the researchers used a time-series dataset containing 14 independent features and one dependent feature. The dataset contained data for every 15 minutes from January 1 to December 31, 2022. “We tested the trained model on another dataset collected from a solar power plant in Durgapur, West Bengal. We then tested the model on a public dataset collected from Denmark. We found that our model gave similar results.”

No model is perfect, and “we faced some limitations during testing,” Dr Giri said, noting that rapid changes in weather parameters gave slightly different results.

When asked whether the ensemble model would require a lot of computing power, Dr Giri said the model is “very lightweight” with just 1.2 million parameters. “We don't think this will be an issue for large-scale implementation,” he said.

“We believe that our models are trained on a very small amount of data,” he said, adding that they are looking to extend their research with larger amounts of data to improve the efficiency of the models.

This work will help researchers explore other dimensions of scientific knowledge in a particular area, not just a particular data set, Dr Giri said. quantum.

Dr. Chandan Giri's full emailed response to Quantum's questions:

How did you train the BI-LSTM component of your model? What kind of data preprocessing was required?

We used Google Colab platform and Python programming language to train the BI-LSTM model. With the help of TensorFlow, scikit-learn, and Numpy libraries, we built and trained the ensemble using Keras API.

Our model has an input layer, two BI-LSTM layers, an average layer, and a final output. Our model receives input with the input shape of the dataset. The two BI-LSTM layers have different initial weights but are trained with the same input. A dropout layer follows the first bidirectional LSTM module in each layer. Again, the bidirectional LSTM module takes the first module output as input and passes through a dropout layer, the output of this layer is fed to another bidirectional LSTM module, and finally the output passes through a dense layer. Similarly, in our other work, the final output is generated by averaging the outputs of both layers. We used LeakyReLu in the LSTM modules and ReLU activation function in the dense layer.

As mentioned in the paper, we first combined the data collected from IIEST Solar Hub with the data collected from CPCBI, and then augmented the dataset with the physical characteristics of the solar panels used in IIEST Solar Hub. During the process of data preprocessing, we first dealt with the missing values ​​in the data, and then dealt with the outliers in the dataset to make the dataset trainable and scaled the independent features.

Can you elaborate on the specific metrics you used to compare your model's performance with traditional models?

To compare the results of the proposed model with the traditional models, a rigorous comparative analysis was conducted using standard performance metrics, namely MSE (Mean Squared Error), RMSE (Root Mean Squared Error), MAE (Mean Absolute Error) and R2 score (coefficient of determination).

How did you ensure that the comparison between your model and other standard deep learning models was fair?

This comparison is fair enough, since we followed the standard rules for comparing performance: we used the same hyperparameters and the same dataset for the proposed model and the other modes.

What was the size and nature of the datasets used for training and evaluation? How did you deal with potential biases in the data?

We used a time series dataset with 14 independent features and one dependent feature. The dataset contains data with 15-minute intervals from January 1, 2022 to December 31, 2022, and includes an 11-hour window (6 AM to 5 PM). This is because most of the data outside this window is close to zero due to very low solar irradiance.

Can you provide more details on how you validated the model's performance across different geographic locations?

We tested our trained model using another dataset collected from a solar power plant at Durgapur WB, India, 150 km away from our laboratory. Meanwhile, we tested our model using a public dataset collected from Denmark and found that our model gives similar results.

Did you find any limitations or weaknesses in your model during testing? How do you plan to address these in future work? Are there specific conditions or types of data where your model does not perform as well as you would expect?

To be honest, there are few or no perfect models. In our work, we faced some limitations during testing. If the weather parameters suddenly changed, we got slightly different results, but the overall performance was very good.

We believe that our models are trained on a very small amount of data, and we are looking to extend our work with larger amounts of data to improve the efficiency of our models, and we are also exploring other possible ways to make our models more robust.

How do you plan to address the computational requirements of training and running ensemble models based on BI-LSTM, especially for large-scale applications?

Our model is very lightweight, with only 1.2 million parameters, so we believe that large-scale implementations should not pose any problems and warrant further investigation.

What do you think is the most novel aspect of this model compared to existing deep learning approaches?

While most of the deep learning models available for solar energy forecasting typically only use meteorological parameters to predict solar power production, we trained our model with some additional features generated using the physical properties of solar panels that make us different from state-of-the-art models.

How do you think your model will impact future research in the areas of time series forecasting and deep learning?

The first author of the study, JRF from the Department of IT at the Indian Institute of Science and Technology, Shibpur, is in the early stages of his PhD studies under the supervision of Dr Chandan Giri, and the study was published two months ago. It has received positive responses from other researchers in India and abroad. We believe that this work will help researchers explore other aspects of the scientific knowledge behind not just a particular dataset but a particular field.

This is your last free article.



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *