Authors | Nantian Huang, Shengyuan Wang, Rijun Wang, Guowei Cai, Yang Liu, Qianbin Dai |
Title | Gated spatial-temporal graph neural network based short-term load forecasting for wide-area multiple buses |
Publication | International Journal of Electrical Power & Energy Systems |
Volume | 145 |
Issue | x |
Pages | x |
Year | 2023 |
DOI | https://doi.org/10.1016/j.ijepes.2022.108651 |
Introduction
Background
- Modern power systems are complex and diverse, incorporating intelligence, flexibility, and networking.
Previous Research
- Bus load forecasting is important for the safety of grid dispatch operations and online analysis decisions in modern power systems.
- Bus load is volatile and random and changes in trends are not obvious, making accurate forecasting difficult.
- Short-term load forecasting methods are divided into statistical and machine learning methods, with machine learning-based methods being more effective in capturing complex nonlinear data features.
- Load forecasting accuracy can be improved by considering spatial-temporal coupling correlation.
- Most existing studies on load forecasting are limited by the sequential input of data in Euclidean space, resulting in insufficient expression of spatial-temporal correlation between different data.
- Graph Convolutional Network (GCN) provides a new technical route for diverse representation and feature mining of data within power systems and has been used in many areas of the power system.
Proposed Model
- To accurately describe the complex spatial-temporal coupling relationship between the wide-area multiple bus loads, the coupling relationship between non-geographic distance and non-grid topology-constrained bus loads is mapped to a similar-weighted spatial-temporal graph.
- A short-term load forecasting method for wide-area multiple buses based on ST-GCN is constructed, expanding the node feature types in the similar-weighted spatial-temporal graph to improve the accuracy of the worst evaluation metrics.
- The RapidMIC value determines node connection relationship and is used as the edge feature in the similar-weighted spatial-temporal graph.
Significance
- A unified model is necessary for accurate short-term load forecasting of multiple bus loads over a wide area.
- Load forecasting accuracy can be improved by analyzing and mining multiple bus loads from non-Euclidean domains.
- ST-GCN can effectively reduce the number of outlier points in the evaluation metrics that considerably affect scheduling in bus load forecasting.
- The constructed similar-weighted spatial-temporal graph can directly reflect the spatial coupling and correlation degree between buses, which further increases the forecasting accuracy of the method.
Proposed Model
- The proposed model is ST-GCN based wide-area multiple bus loads forecasting method.
- The method constructs SCL to extract node features within the similar-weighted spatial-temporal graph.
- SCL focuses on node features within a K-order neighborhood centered on a node and is abstracted as the feature transfer of each node.
- The extensive node features in the similar-weighted spatial–temporal graph are extracted by stacking multiple layers of the SCL.
- GRUL with a gating mechanism is built to mine temporal domain features between similar-weighted spatial-temporal graph.
- The high-dimensional feature vector output from the SCL is used as the input to the GRUL to achieve a wide-area multiple bus loads forecast.
- The GRUL outputs ht using the features extracted by SCL with the hidden layer output ht-1 at the moment t-1 as the input.
- The proposed method is not limited by the sequential input of data and can fully consider the influence of spatial-temporal correlation between bus loads on the forecasting results.
Experiment
- RapidMIC method identifies nodal meteorological features and input features.
- Temperature is identified as a strongly correlated feature with load, with a RapidMIC of 0.51.
- Model input features are consistent with nodal meteorological features for all experiments.
- ST-GCN method is compared with CNN-LSTM, GRU, SVR, and ST-GCN* to carry out comparison experiments.
- ST-GCN has a significantly lower box structure in SMAPE metrics compared to CNN-LSTM, GRU, and SVR methods.
- The ST-GCN method has fewer and lower value outliers and can effectively improve load forecasting accuracy after accounting for complex multivariate features.
- The worst evaluation metrics in the test set are improved by 1.82 % to 5.94 %.
댓글