This Issue Selection
2024 Issue 9
Wang Yongsheng, Li Hailong, Guan Shijie, Wen Caifeng, Xu Zhiwei, Gao Jing
DOI: 10.13336/j.1003-6520.hve.20231942
According to the “2023 Global Wind Power Report,” as of the end of 2022, the global installed wind power capacity reached 906 GW, an increase of 9% compared to 2021. The rapid growth in the proportion of wind power not only promotes the clean and low-carbon development of energy but also effectively safeguards national energy security. However, wind power generation has characteristics such as intermittency, high variability, and strong randomness, which pose significant challenges and adverse impacts on grid integration. Therefore, accurately predicting wind power output has become one of the key strategies to address the grid integration issues of wind power.
1) Establishment of the Combined Model Based on Transform Domain Analysis and XGBoost Algorithm
To achieve accurate and reliable wind power prediction, this study proposes a short-term wind power prediction method based on transform domain analysis and the XGBoost algorithm (XGBoost transformer domain analysis, XGBoost-TDA). Figure 1 clearly illustrates the framework of the XGBoost-TDA model, which mainly includes two key modules: one is a multi-level feature learning module based on wind power indicators and transform domain analysis; the other is a prediction module based on the XGBoost algorithm.
Figure 1 Short-Term Wind Power Prediction Model Based on Transform Domain Analysis and XGBoost Algorithm
The construction of the XGBoost-TDA model for wind power prediction involves five steps, with the main tasks of each step as follows:
1) Obtain wind power data, check the integrity of the original dataset, and handle outliers and missing values.
2) Construct a time window to prepare for feature engineering calculations.
3) Calculate wind power indicators, FFT, and Haar wavelet transform, and synthesize new data with usable covariates.
4) Train the model based on the partitioned training set and use the test set to validate the model’s predictive performance.
5) Report the prediction results to the wind farm to assist in adjusting generation plans and residential electricity usage plans.
2) Evaluation of Model Prediction Results
To demonstrate the predictive effect of the model proposed in this study, we take the Dataset 1 Inner Mongolia Biluohe Wind Farm dataset and Dataset 2 Tennet Wind Power Generation dataset from Germany as examples. The prediction results of wind power, including actual power values, predicted values from the proposed model, and six sets of comparative model predicted values, are fully displayed, as shown in Figure 2. Among them, the solid line represents the actual values, the dashed line represents the predicted values from the XGBoost-TDA model, and other differently styled lines represent the predicted values from comparative models.
Figure 2 Comparison of Prediction Results of XGBoost-TDA and Other Models on Different Datasets
From Figure 2, it is evident that the XGBoost-TDA model proposed in this study exhibits a high degree of consistency between the predicted values and actual values compared to other models, demonstrating superior performance. Specifically, the XGBoost-TDA model shows the best consistency when processing Dataset 2, with its predicted values being very close to the actual values, forming a good fitting curve. Although there are some discrepancies between the predicted values and actual values in Dataset 1, it still achieves the best correspondence compared to other models. This indicates that by adopting transform domain analysis methods and the XGBoost algorithm, the predictive results of the model proposed in this study are closer to the actual values.
1) Combining the transform domain multi-level feature learning framework with fast Fourier transform and Haar wavelet transform, each model component is designed to maximize its advantages, allowing for better extraction of potential information between data from complex wind power-related time series, uncovering the inherent relationships between the data, thereby significantly improving the model’s accuracy and generalization ability.
2) By applying the XGBoost model, training time is reduced, enhancing the timeliness of applications in practical production environments, thus further improving the predictive efficiency of the proposed model, achieving faster and more accurate wind power predictions. The proposed model requires only 0.941 seconds to complete training, making the lightweight prediction algorithm feasible for deploying the model to edge computing devices in wind farms for prediction.
3) By combining multi-level transform domain feature engineering with machine learning models, predictive results comparable to or even surpassing deep learning are achieved. For instance, in terms of ENMAE, the proposed model reduces about 2.18% compared to baseline models.
4) The effectiveness and applicability of the proposed model are validated using real datasets from multiple wind farms, confirming the robustness of the model.
The method proposed in this paper effectively resolves meaningless feature extraction methods, fully considering the importance of frequency domain information in feature learning, and achieves fast and accurate short-term wind power predictions. This aids power departments in adjusting generation plans, formulating scheduling schemes, and provides effective references for other time series prediction scenarios (such as wind speed, photovoltaic generation, and precipitation).
To further enhance the prediction performance of the model for practical applications in wind farms, future research plans to explore more methods for feature extraction. For example, introducing graphical convolutional neural networks into the model to explore the spatial-temporal relationships of potential data. Additionally, heuristic algorithms for dynamic adjustment can be added, focusing on setting the optimal parameters for the model to further enhance the performance of the prediction model.
Citation Information:
Wang Yongsheng, Li Hailong, Guan Shijie et al. Short-Term Wind Power Prediction Model Based on Transform Domain Analysis and XGBoost Algorithm[J]. High Voltage Technology, 2024, 50(9): 3860-3870.
Wang Yongsheng, PhD, Professor, Senior Engineer, focuses on research in the field of energy, especially renewable energy security, using data analysis and mining technologies, machine learning, and deep learning for energy time series data analysis and mining. His research results are widely applied in wind and solar power plant production scheduling, IT, and energy system operation and maintenance. He has led over 10 provincial and ministerial scientific research projects, guided more than 10 master’s students, and published 12 relevant papers as the corresponding author and first author in SCI/EI/Baidu core journals and CCF listed academic conferences, with 10 patents granted (applied).
Li Hailong, Male, Master’s student, mainly engaged in research on energy security, big data processing, and wind power prediction.
Guan Shijie, Male, PhD student, mainly engaged in research on renewable energy generation power prediction, integrated energy system modeling, and energy flow optimization. He has published 5 academic papers as the first author in well-known journals at home and abroad, with a total of 41 citations.
Editor: He Qiuping
– 【 END 】 –
Disseminating Knowledge · Sharing Achievements · Exchanging Experiences
Statement: This article is an original work, and all text and image copyrights belong to the Editorial Department of “High Voltage Technology.” According to the latest regulations of the National Copyright Administration, for reprinting or excerpting our works in print media, websites, or WeChat public accounts, please contact our editorial department in advance. Individuals are allowed to forward and share this WeChat original text.
#Hot Articles, #Hot Reports, #Call for Papers, #Conference Information, #Online First Release, #Selected Papers
http://hve.epri.sgcc.com.cn/
“High Voltage Technology” Video Account
Contact Us
