Abstract
The power dispatching network forms the backbone of efforts to automate and modernize power grid dispatching, rendering it an indispensable infrastructure element within the power system. However, accurately forecasting future flows remains a formidable challenge due to the network’s intricate nature, variability, and extended periods of missing data resulting from equipment maintenance and anomalies. Vital to enhancing prediction precision is the interpolation of missing values aligned with the data distribution across other time points, facilitating the effective capture of nonlinear patterns within historical flow sequences. To address this, we propose a transfer learning approach leveraging the gated recurrent unit (GRU) for interpolating missing values within the power dispatching network’s flow sequence. Subsequently, we decompose the generation of future flow predictions into two stages: first, extracting historical features using the GRU, and then generating robust predictions via eXtreme Gradient Boosting (XGBoost). This integrated process termed the GRU-XGBoost module, is applied in experiments on four flow sequences obtained from a power grid company in southern China. Our experimental findings illustrate that the proposed flow prediction model outperforms both machine learning and neural network models, underscoring its superiority in short-term flow prediction for power-dispatching networks.
Introduction
Power dispatching networks serve as the "nervous system" of the modern electrical grid, coordinating generation and consumption to ensure stability. Effective management requires accurate short-term flow predictions to avoid congestion and optimize resources. However, researchers face two primary hurdles:
- Missing Data: Equipment maintenance and sensor anomalies frequently cause long gaps in data. Standard interpolation methods often fail to capture the complex, non-linear nature of these power flows.
- Feature Complexity: Power flow data is highly volatile. While neural networks excel at sequence modeling, they can struggle with final regression tasks compared to robust gradient-boosting algorithms.
Methodology
The proposed approach addresses these challenges through a two-stage framework involving data imputation and a hybrid prediction module.
1. GRU-Based Transfer Learning for Imputation
To handle missing data, a transfer learning strategy is employed. A Gated Recurrent Unit (GRU) model is first trained on a "source domain"—a highly correlated, complete flow sequence. This model is then fine-tuned on the "target domain" (the sequence with missing values). This allows the system to intelligently "fill in the blanks" based on learned patterns from similar sensors, significantly improving data quality over simple linear interpolation.
2. The GRU-XGBoost Prediction Module
Once the data is cleaned, the prediction task is split into two specialized roles:
- Feature Extraction (GRU): The GRU processes historical sequences to reconstruct and encode temporal features into a hidden state vector.
- Robust Prediction (XGBoost): The encoded features are fed into eXtreme Gradient Boosting (XGBoost). As an ensemble of decision trees, XGBoost is exceptionally effective at handling non-linear relationships and preventing overfitting, providing a more robust final output than standard neural network layers.
Experimental Setup and Results
The model was validated using real-world data from a power grid company in Southern China, spanning December 2020 to April 2021. The dataset included four distinct flow sequences (x_in, x_out, s_in, and s_out) representing different network regions.
The hybrid GRU-XGBoost model consistently outperformed traditional models like MLP, SVR, and even the Transformer-based Informer model. For the x_out dataset, the model achieved:
- MAE (Mean Absolute Error): 1.86E+07
- MAPE (Mean Absolute Percentage Error): 2.71% (outperforming MLP at 2.88% and SVR at 3.24%)
- RMSE (Root Mean Square Error): 2.48E+07
Conclusion
By decoupling feature extraction from regression and utilizing transfer learning for data imputation, the GRU-XGBoost framework offers a significant advancement in power grid flow forecasting. While the model requires more computational resources and careful hyperparameter tuning, its ability to handle missing data and non-linear patterns makes it a superior choice for modern power infrastructure management. Future work will focus on improving generalizability and reducing the impact of random noise.
Summary updated