Leveraging Spatial Information in Smart Grids using STGCN for Short-Term Load Forecasting
Leveraging Spatial Information in Smart Grids using STGCN for Short-Term Load Forecasting
复制标题
DOI：
10.1145/3474124.3474145
复制
发表时间：
2021-08
期刊：
影响因子：
--
通讯作者：
C. Cheung;S. Kuppannagari;R. Kannan;V. Prasanna
中科院分区：
文献类型：
--
作者：
C. Cheung;S. Kuppannagari;R. Kannan;V. Prasanna
预测未来几个时段内能源用户(负荷)行为的问题--短期负荷预测(STLF)对几个电网运营的成功至关重要。由于数据的高波动性，很难在较低的聚合水平上进行预测。智能电网的运行，以及由此产生的任何数据，都表现出高度的空间相关性，这是由于配电网的拓扑结构以及其他潜在因素，如邻域的相似性、社会经济状况等。虽然时间信息通常被利用在递归或卷积层等神经网络结构中，但空间信息在负荷预测中的使用尚未被探索。针对智能电网短期负荷预测问题，提出了一种时空图卷积网络(STGCN)模型。STGCN专门捕捉数据中的空间和时间相关性，以获得更准确的预测。我们还表明，我们的模型，通过捕获空间和时间相关性，比最先进的预测模型对缺失数据具有更强的鲁棒性。我们在美国爱荷华州的一个低聚合级别(每个数据点5个∼10个客户)的数据集上进行了详细的评估，结果表明，我们的模型预测了提前3小时的实际负载消耗，平均绝对误差比最佳基准模型低7.54%，如果数据有缺失，则均方根误差最高可降低38.72%。
The problem of predicting the behaviour of energy consumers (loads) in the next few intervals — Short-Term Load Forecasting (STLF) is critical to the success of several grid operations. Prediction at lower aggregation levels is difficult due to the high volatility of the data. Smart grid operations, and in turn any data generated as a result of them, exhibit high spatial correlations imposed due to the topology of the power distribution network as well as other latent factors such as similarity in neighborhood, socio-economic status, etc. While temporal information is usually leveraged in neural network structures like Recurrent or Convolutional Layers, the use of spatial information in load forecasting has not been explored. In this paper, we develop a Spatial-Temporal Graph Convolutional Network (STGCN) model for the problem of Short-Term Load Forecasting in Smart Grids. STGCNs specialize in capturing both spatial and temporal correlations in the data to obtain more accurate predictions. We also show that our model, by capturing both spatial and temporal correlations, is more robust to missing data than state-of-the-art prediction models. We perform detailed evaluation on a dataset based in Iowa, US with real power at a low aggregation level (5 ∼ 10 customers per datapoint) and show that our model predicts 3 hours ahead real load consumption with a Mean Absolute Error of 7.54% less than the best performing baseline model, and as much as 38.72% less in Root Mean Squared Error (RMSE) if the data has missing entries.