Short-Term Prediction of PV Power Based on Combined Modal Decomposition and NARX-LSTM-LightGBM

Hongbo Gao, Shuang Qiu, Jun Fang, Nan Ma, Jiye Wang, Kun Cheng, Hui Wang, Yidong Zhu, Dawei Hu, Hengyu Liu, Jun WangSustainability 2023 · DOI

TL;DR

This paper introduces a novel hybrid machine learning model, CD-NARX-LSTM-LightGBM, for accurate short-term photovoltaic (PV) power prediction. By combining modal decomposition techniques with NARX, LSTM, and LightGBM, the model effectively captures complex non-linear patterns in PV generation, leading to significantly improved prediction accuracy across various weather conditions.

What Problem Does This Paper Address?

The increasing integration of solar energy into power grids presents a challenge due to the inherent variability and non-linearity of PV power generation. Fluctuations caused by weather conditions like clouds and varying sunlight intensity can destabilize the grid, impacting power dispatch and scheduling. Accurate short-term PV power prediction is crucial for mitigating these impacts, ensuring grid stability, and optimizing energy management. Existing prediction models often struggle with the complex, non-stationary nature of PV data, leading to suboptimal accuracy.

Methods

The proposed approach, CD-NARX-LSTM-LightGBM, is a multi-stage hybrid model:

  1. Feature Selection: Pearson correlation analysis is used to identify environmental variables (ambient temperature, irradiance, etc.) that have a strong correlation with PV power.
  2. Combined Modal Decomposition (CD): The original PV power signal is decomposed into multiple intrinsic mode functions (IMFs) using Empirical Mode Decomposition (EMD), Ensemble EMD (EEMD), and Complete Ensemble EMD with Adaptive Noise (CEEMDAN). This process aims to reduce signal complexity and extract underlying patterns.
  3. Feature Matrix Construction: IMFs that exhibit strong correlations with PV power are selected to form a new feature matrix.
  4. NARX for Error Correction: A Nonlinear Auto-Regressive Neural Network with Exogenous Input (NARX) is used to generate preliminary PV power predictions and an "error correction vector." This vector helps capture residual errors from the initial prediction.
  5. NARX-LSTM-LightGBM Integration: The preliminary NARX predictions and the error correction vector are then fed as input to a combined NARX-LSTM-LightGBM model. This hybrid model leverages the strengths of NARX for capturing dynamic interactions, LSTM for handling sequential data, and LightGBM for efficient gradient boosting.
  6. Error Inverse Method: An error inverse method is employed to optimally combine the predictions from different model components, further refining the final output.

Datasets

The study utilizes a real-world dataset from Andre Agassi College, USA. This dataset spans from January 1, 2012, to December 31, 2014, with 15-minute intervals. The data includes seven environmental characteristic variables: ambient temperature, PV inverter temperature, module temperature, irradiance, ambient humidity, wind speed, and wind direction. For model training and testing, data from February 6, 2014, to December 17, 2014, was used. The test set was specifically curated to include six days representing three typical weather types: sunny, cloudy, and rainy.

Key Insights and Ideas

  • Decomposition is Key: Decomposing the complex PV power signal into its constituent IMFs is crucial for revealing hidden patterns and reducing noise, which significantly aids subsequent prediction models.
  • Synergy of Models: Combining different types of models (NARX for dynamics, LSTM for sequences, LightGBM for efficiency) creates a more robust prediction system than any single model alone.
  • Error Correction Matters: Using NARX to generate an "error correction vector" and then incorporating it into the main prediction model helps to refine predictions and account for residual errors, especially in challenging weather conditions.
  • Weather Dependency: The performance of prediction models can vary significantly with weather. The proposed model shows particular strength in improving predictions for cloudy and rainy days, where PV generation is more volatile.

Key Results

The proposed CD-NARX-LSTM-LightGBM model demonstrated superior performance compared to individual models and other combined approaches.

  • Accuracy Improvements: Under various weather conditions, the model achieved Root Mean Square Error (RMSE) lower than 1.665 kW, Mean Absolute Error (MAE) lower than 0.892 kW, and Mean Absolute Percentage Error (MAPE) lower than 0.211%.
  • Specific Reductions: For example, on rainy day 2, the CD-NARX-LSTM-LightGBM model showed a reduction of 57.19% in RMSE, 25.43% in MAE, and 25.34% in MAPE compared to the NARX-LSTM-LightGBM model.
  • Robustness: The model exhibited higher prediction stability across different weather types, with RMSE values consistently within a narrow range.
  • Computational Efficiency: While the combined modal decomposition adds to the processing time (average run time of 67.59 min for CD-NARX-LSTM-LightGBM vs. 45.23 min for NARX-LSTM-LightGBM), the gains in accuracy are substantial.

Limitations and Future Work

  • Computational Cost: The modal decomposition step, while beneficial for accuracy, increases the overall computation time.
  • Future Work: The authors suggest expanding the scope of their models to include predictions for other smart grid applications, such as wind power and load forecasting.

Summary updated

Abstract

Recently, solar energy has been gaining attention as one of the best promising renewable energy sources. Accurate PV power prediction models can solve the impact on the power system due to the non-linearity and randomness of PV power generation and play a crucial role in the operation and scheduling of power plants. This paper proposes a novel machine learning network framework to predict short-term PV power in a time-series manner. The combination of nonlinear auto-regressive neural networks with exogenous input (NARX), long short term memory (LSTM) neural network, and light gradient boosting machine (LightGBM) prediction model (NARX-LSTM-LightGBM) was constructed based on the combined modal decomposition. Specifically, this paper uses a dataset that includes ambient temperature, irradiance, inverter temperature, module temperature, etc. Firstly, the feature variables with high correlation effects on PV power were selected by Pearson correlation analysis. Furthermore, the PV power is decomposed into a new feature matrix by (EMD), (EEMD) and (CEEMDAN), i.e., the combination decomposition (CD), which deeply explores the intrinsic connection of PV power historical series information and reduces the non-smoothness of PV power. Finally, preliminary PV power prediction values and error correction vector are obtained by NARX prediction. Both are embedded into the NARX-LSTM-LightGBM model pair for PV power prediction, and then the error inverse method is used for weighted optimization to improve the accuracy of the PV power prediction. The experiments were conducted with the measured data from Andre Agassi College, USA, and the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) of the model under different weather conditions were lower than 1.665 kw, 0.892 kw and 0.211, respectively, which are better than the prediction results of other models and proved the effectiveness of the model.
BibTeX
@inproceedings{gao2023short,
  author = {Hongbo Gao and Shuang Qiu and Jun Fang and Nan Ma and Jiye Wang and Kun Cheng and Hui Wang and Yidong Zhu and Dawei Hu and Hengyu Liu and Jun Wang and Jun Wang and Jun Wang},
  title = {{Short-Term Prediction of PV Power Based on Combined Modal Decomposition and NARX-LSTM-LightGBM}},
  booktitle = {Sustainability},
  year = {2023},
  doi = {10.3390/su15108266},
}
Figure