Predicting the Power Output of Solar Panels based on Weather and Air Pollution Features using Machine Learning

Abstract

The power output of solar panels highly depends on environmental situations like weather and air pollution. Due to bad weather or air pollution, it is difficult for solar panels to operate at their full potential. Knowing the power output of solar panels in advance helps set up the solar panels correctly and work their possible potential. This paper presents an approach to predict the power output of solar panels based on weather and air pollution features using machine learning methods. We create machine learning models with three kinds set of features, such as weather, air pollution, and weather and air pollution. Our datasets are collected from the area of Seoul, South Korea, between 2017 and 2019. The experimental results show that the weather and air pollution features can be efficient factors to predict the power output of solar panels.

1. INTRODUCTION

Solar power is one of the efficient energy sources that is captured from the sun using solar panels. South Korea is an active investor in solar panel installations, which ranks them in the top ten markets of solar energy in the world[1]. South Korea contributed significantly to 3.1 GW (gigawatt) power in 2019[1]. With the increasing interest in solar power[2], it is a challenge to use its full potential. The power output of solar panels highly depends on environmental situations like weather and air pollution. Due to bad weather or air pollution, it is difficult for solar panels to operate at their full potential. Thus, knowing the power output of solar panels in advance could help correctly set up the solar panels and make sure that these panels operate at their full potential.

In this paper, we argue that the environmental situation is important to consider for producing power using solar panels. It is obvious that solar panels can operate at their full potential when the sun shines and there is no partial shading on them. However, the partial shading might occur due to temporary effects, such as weather (e.g., clouds or rain) and air pollution (e.g., fine dust). This paper presents an approach to predict the power output of solar panels based on weather and air pollution features using machine learning methods. We aim to reveal the best combination of environmental features and machine learning methods. We collect datasets between 2017 and 2019 from Seoul area, South Korea.

This paper builds power output prediction models using the state-of-art machine learning models, such as Linear Regression, K-Nearest Neighbors (kNN), Support Vector Regression (SVR), MultiLayer Perceptron (MLP), Random Forest Regressor (RF), and Gradient Boosting Regressor (GB). Each model uses three sets of features. The first set includes weather features, such as humidity, sunshine, solar radiation, and clouds. The second set consists of air pollution features, which include Ozone (O₃), Sulfurous Acid Gas (SO₂), Nitrogen Dioxide (NO₂), Carbon Monoxide (CO), Fine Dust (PM₁₀), and Fine Particulate Matter (PM_2.5). Finally, the third set is created by choosing top correlated features from the other two sets. We evaluated the models using popular error rate methods, such as Mean Absolute Error (MAE), Coefficient of Determination (R²), and Root Mean Squared Error (RMSE). The experimental results show that the weather and air pollution features can be efficient factors to predict the power output of solar panels.

We organized the rest of this paper as follows. Section 2 includes the brief descriptions of similar studies related to our paper. Section 3 describes the methodology of our paper. Section 4 explains the experimental works. Finally, Section 5 discusses the conclusion and determines future works.

2. RELATED STUDY

As we mentioned in Section 1, environmental situations like weather and air pollution are features that must be considered when predicting the power output of solar panels [2-9]. For example, several researchers proposed the prediction models to estimate the power output of solar panels using weather features[3-6]. Nageem et al.[3] proposed a multi-input support vector regression (SVR) model to predict the output of so-called a grid-connected solar panel. The authors used weather features like humidity, temperature, pressure, and wind speed. The analytical and SVR models were compared in this paper, and the experiment results show that the SVR model had more accuracy than the analytical model. Saberian et al.[4] used artificial neural networks (ANNs) to predict the output power of solar panels based on weather features like temperature and irradiance. The authors used a dataset of five years from 2006 to 2010. The experiments show that the proposed model achieves strong performance.

Besides, there are many approaches to predict the output power of solar panels using air pollution features[7-9]. Son et al.[7] studied the effect of particulate matter (PM) on solar power output over South Korea. The authors used the PM_2.5 and PM₁₀ concentrations for the dataset from 2015 to 2017. The authors mentioned that the PMs commonly reduce solar power output by more than 10%. Their results show that PM's negative impacts on solar panels should be considered in policymaking on target solar power output in South Korea. Bergin et al.[8] calculated the solar energy production reduction due to dust and particulate air pollution. The authors merged field measurements and global modeling to evaluate the influence of dust and PM on solar electricity generation. The result indicated that solar panel production was reduced by 17 - 25% due to PM on the surface of solar panel

3. METHODOLOGY

3.1 Overview

Fig. 1 describes the overall flow of our methodology. It consists of the following steps: data acquisition, feature extraction, model training, and prediction of power output. First, our experimental datasets are created from three different sources that were mentioned in Section 1. After that, we build machine learning models by selecting different sets of features. Here, we aim to reveal the best combination of features and machine learning methods. All steps are explained in the following subsections.

Fig. 1. The overall flow of the methodology.

3.2 Datasets

There are three data sources, where we collect our experimental datasets. The power output data of solar panels are provided by a South Korean renewable energy power output company called Daeyeon C&I. Further, the weather-related data is collected from South Korea Meteorological (KMA). Lastly, we collected air pollution data from the Seoul Metropolitan Government (SMG). All datasets are generated in the location of Seoul, South Korea. Table 1 shows the feature description of collected data. From the table, we can perceive that the power output is our target feature. Other features are inputs to our models, and are divided into four sections like solar panel. weather, air pollution, and derived features. The derived features are generated from the date feature that power output is generated. We combined all datasets based on the date. The dataset contains 14191 observations from 2017 to 2019. We split the dataset by 80% and 20% as train and test, respectively. Our models predict next one-hour power output.

Table 1. The feature description of datasets.

3.3 Feature Extraction

After combining the separated datasets, the first step is to obtain the features appropriate to predict power output solar panels. First, we extend each set of features using derived features (month, hour), which are generated from date that the power output is registered. Fig. 2 shows the relationship between power output of solar panel, month and hour. From the Fig. 2 (a), we can perceive that the power output of solar panels is produced more in April to June. We can assume that it is the time when there is a plenty of sunlight, and the level rainfall is low. On the contrary, Fig. 2(b) shows the power output of solar panels by hour. From 10 to 13 hours is the rush hour to generate power from solar panels.

Fig. 2. Correlation between power output and derived features.

Further, we determine linear relationship between two features using Pearson’s correlation. The result of correlation between -1 and 1; here -1 represents a negative relation, 1 represents a positive relation. If a result is near to zero, it means there is no relation between features. Fig. 3 shows the correlation between power output and other features. Based on the result showed in Fig. 3, we created five sets of features. First set is solar panel features, second set is weather features, and third set is air pollution features. Also, fourth set is created by selecting top correlated features among the previous three sets, and fifth set is created by including all features. In other words, we essemble solar panel features (i.e., month, hour, power factor, slope, horizontal irradiation, module temperature), weather features (i.e., month, hour, humidity, sunshine, solar radiation, clouds, temperature), air pollution features (i.e., month, hour, O₃,SO₂,NO₂,CO,PM₁₀,PM_2.5), top correlated features (i.e., power factor, slope, horizontal irradiation, module temperature, sunshine, solar radiation), and all features.

Fig. 3. Pearson’s correlation between power output and other features.

3.3 Machine Learning Models

This paper builds power output prediction models using state-of-art machine learning models, such as Linear Regression, SVR, MLP, kNN, RF, GB. All models are written in Python and are created using Scikit-learn library[10]. Linear regression is one of the well-known statistical and machine learning algorithms. Linear regression calculates output variable (y) by input variables (x) using least-squares to fit a line to the data points. SVR is a regression version of Support Vector Machines (SVM), which is like Linear Regression in defining the hyperplane in data points. The difference is that SVR uses a threshold to minimize the error between actual and forecasted values. MLP is a field of feedforward artificial neural networks, consists of an input, hidden, and output layers. The hidden and output layers are neurons that use a nonlinear activation function. kNN regression is a machine learning method that predicts value by averaging the observations in the same neighbor. Choosing an appropriate number of neighbors is the main object of the kNN method. Random Forest (RF) is one of popular supervised machine learning algorithms for classification and regression. RF can work well with a dataset that has a large number of features. In addition, it indicates the importance of variables. However, RF is not well adapted for categorical data. Gradient Boosting (GB) is a powerful algorithm in the area of classification and regression problems. Here, boosting combines multiple simple models into a single complex model. Therefore, GB uses the decision tree as a single model. Table 2 compares the pros and cons of selected machine learning methods.

Table 2. The pros and cons of machine learning methods.

4. EXPERIMENTAL RESULTS

4.1 Metrics of Evaluation

To evaluate the performance of models, we compared the actual power output of solar panels with the predicted power outputs according to certain error rate methods. Specifically, we use R², RMSE and MAE. RMSE is the standard deviation of the prediction errors. MAE calculates the average of differences between the actual value and predicted value. The lower values indicate better accuracy in the RMSE and MAE. R² is also a regression score method. Best R² score is 1 and it can be negative value, which expresses worser accuracy. We multiply the R² score to represent it by percentage. R², RMSE, MAE are given in Equations (1)–(3), respectively. Here, N represents the number of samples, y represents an actual value, ŷ represents a predicted value, \(\bar{y}\) represents the mean of actual values.

\(R^{2}=\left(1-\frac{\sum_{i=1}^{N}\left(y_{i}-\hat{y_{i}}\right)^{2}}{\sum_{i=1}^{N}\left(y_{i}-\bar{y}\right)^{2}}\right)* 100\)       (1)

\(\mathrm{RMSE}=\sqrt{\frac{1}{\mathrm{~N}} \sum_{\mathrm{i}=1}^{\mathrm{N}}\left(\mathrm{y}_{\mathrm{i}}-\hat{\mathrm{y}_{\mathrm{i}}}\right)^{2}}\)       (2)

\(M A E=\frac{1}{N} \sum_{i=1}^{N}\left|y_{i}-\hat{y_{i}}\right|\)       (3)

4.2 Results

To evaluate the performance of models, Finding the correct values of hyperparameters is an important step to build an efficient model, especially for a parameter-sensitive algorithm like SVR. We performed each model with several combinations of hyperparameter values to find out the best option. Table 3 shows the values of the hyperparameters of each model.

Table 3. The parameter settings of machine learning models.

Table 4 describes detailed results of all models and sets of features. From the table, we can see that RF, GB, and MLP shows better accuracy than kNN, SVR, and Linear Regression. Also, solar panel and weather features are more efficient than air pollution features. Solar panel features show best performances with all models because slop and horizontal irradiations are highly correlated with power output of solar panels. On the contrary, there are no significant differences between weather features and top-correlated features. It means that weather features are appropriate enough to predict the power output of solar panels. But air pollution features show good accuracy with RF and GB, with around 67% of R^2. RF is the most appropriate model that shows best accuracy with all sets of features. The best R², RMSE, MAE of RF are 98.25%, 0.89, and 0.28, respectively. The range power output of solar panels is 0 to 25.

Table 4. The result of machine learning models.

Fig. 4 shows the RMSE of all models with three sets of features to show the accuracy difference of each model. The results of RF, GB, MLP are alike to each other, and kNN, SVR, and Linear Regression results are similar. RF and GB have the best accuracy and there are no big differences between them. The result of RF model is around 0.8% lower than GB model, 42.6% lower than Linear Regression, 27.9% lower than SVR, 6.6% lower than MLP, and 11.9% lower than kNN models.

Fig. 4 RMSE of machine learning methods.

Fig. 5 displays the power output prediction of solar panels with (a) solar panel features, (b) weather features, (c) air pollution features, (d) top correlated features, and (e) all features. The power output prediction of solar panels is the result of the RF model because the RF model has the best accuracy than other machine learning models. Here, the blue line represents real value, and the yellow line represents predicted value. The figure demonstrates how the predicted values follows the real values. Similar to the above explanations of experiments in Table 4 and Fig. 4, the predictions of weather features show similar values to real values than air pollution feature.

Fig. 5. The power output of solar panels prediction. The result of RF Model.

5. CONCLUSION

This paper have proposed to predict the power output of solar panels based on weather and air pollution features. Here, five sets (i.e., solar panel, weather, air pollution, top correlated, all features) of features have been experimented using six kinds of machine learning methods, such as Linear Regression, SVR, MLP, kNN, RF, and GB. We demonstrated that solar panel features have achieved best accuracy of around 95% in all models. Weather and top correlated features have shown shown the similar good results. However, we have also learned that the air pollution features did not show the good results compared to two other experiments. Lastly, we have created the best model in the the form of RF model with around 98% accuracy.

Our paper aims to reveal appropriate features to predict the power output of solar panels. Solar panels can gather energy well when the sunlight reflects on the solar panels without any obstacle. However, there are many times when the sunlight cannot reflect directly on the solar panels due to environmental situations like weather or air pollution. Thus, the correct placement of solar panels is also depended on these environmental situations. Also, solar panel users can know the time to clean solar panels based on power output prediction. For example, the prediction system can notify users that the solar panels will produce powerless in the future. Because the solar panels can be covered with snow, rain, or air pollution elements. In general, predicting the power output of solar panels in advance can bring various benefits, including adjusting installation, monitoring operation, and planning the business strategies in the future.

Furthermore, we can add more environmental features like wind, temperature, and others to increase the accuracy of models. Also, due to irregular solar panels operation, there can be possibility of outliers in the power output data

points. Thus, we can consider removing outliers that may help further increase the accuracy of prediction. Finally, deep learning models like LSTM can be applied in the future to increase the accuracy of power output prediction even further.

[script] var kistiSpellChecker_replaceCKEditorOptionId = setInterval(function() { if ( ! window.CKEDITOR ) return; clearInterval(kistiSpellChecker_replaceCKEditorOptionId); var isMarkTagAllowed = function(config) { if (config.allowedContent === true) return true; if (typeof(config.allowedContent) === 'string' && config.allowedContent.includes('mark') ) { return true; } if (typeof(config.allowedContent) === 'object' && config.allowedContent.keys().includes('mark') ) { return true; } if (typeof(config.extraAllowedContent) === 'string' && config.extraAllowedContent.includes('mark') ) { return true; } if (typeof(config.extraAllowedContent) === 'object' && config.extraAllowedContent.keys().includes('mark') ) { return true; } return false; }; for (var key in CKEDITOR.instances) { var editor = CKEDITOR.instances[key]; var element = editor.element; var config = editor.config; if ( isMarkTagAllowed(config) ) { continue; } config.extraAllowedContent = 'mark; ' + (config.extraAllowedContent || ''); var originEditorKey = null; for (var key in window) { if ( window[key] === editor ) { originEditorKey = key; break; } } var extraProperties = { listSeq: editor.listSeq // KISTI 기계학습데이터 구축 작업화면 대응 }; console.log('listSeq before destroy', editor.listSeq); editor.destroy(); window[originEditorKey] = editor = CKEDITOR.replace(element, config); for (var key in extraProperties) { editor[key] = extraProperties[key]; } console.log('extraProperties after recreate', extraProperties); } }, 100);[/script] [script] var kistiSpellChecker_replaceCKEditorOptionId = setInterval(function() { if ( ! window.CKEDITOR ) return; clearInterval(kistiSpellChecker_replaceCKEditorOptionId); var isMarkTagAllowed = function(config) { if (config.allowedContent === true) return true; if (typeof(config.allowedContent) === 'string' && config.allowedContent.includes('mark') ) { return true; } if (typeof(config.allowedContent) === 'object' && config.allowedContent.keys().includes('mark') ) { return true; } if (typeof(config.extraAllowedContent) === 'string' && config.extraAllowedContent.includes('mark') ) { return true; } if (typeof(config.extraAllowedContent) === 'object' && config.extraAllowedContent.keys().includes('mark') ) { return true; } return false; }; for (var key in CKEDITOR.instances) { var editor = CKEDITOR.instances[key]; var element = editor.element; var config = editor.config; if ( isMarkTagAllowed(config) ) { continue; } config.extraAllowedContent = 'mark; ' + (config.extraAllowedContent || ''); var originEditorKey = null; for (var key in window) { if ( window[key] === editor ) { originEditorKey = key; break; } } var extraProperties = { listSeq: editor.listSeq // KISTI 기계학습데이터 구축 작업화면 대응 }; console.log('listSeq before destroy', editor.listSeq); editor.destroy(); window[originEditorKey] = editor = CKEDITOR.replace(element, config); for (var key in extraProperties) { editor[key] = extraProperties[key]; } console.log('extraProperties after recreate', extraProperties); } }, 100);[/script]