Research On Electric Power Big Data Technology For Extended Short-term Load Ensemble Forecasting Of Distribution Network

Short-term load forecasting in distribution networks is essential for the safety,stability,and economic viability of power systems.Forecasting load in the short-term is also instrumental in the development and transformation of power grids.Extended short-term load forecasting is often referred to as predicting grid load for shorter tie-scales than conventional short-term load forecasting of a day,e.g.,hour,or dozens of minutes.Both conventional short-term load forecasting and extended short-term load forecasting belong to forecasting load in the short-term,with basically the same principles.Compared with conventional short-term load forecasting,the requirements for quasi-real-time data management and data analysis and mining are stricter for extended short-term.In this study,a two-tier electric power big data architecture is proposed based on the abstract modeling and analysis of the formation process of the short-term load forecasting capability.The tiers include data storage management,and data analysis and mining tier.Electric power big data technologies for improving quasi-real-time data storage management and enhancing supervised learning quality on sample load datasets are further devised to support the development of short-term load forecasting.In the proposed method,the inverted secondary index cluster of the distributed No SQL database is obtained based on the cluster framework.The inverted index data structure is then considered as the main body to research the technology in the data storage management tier.Also,by modeling the time-consuming of cluster-based retrieval and dissecting the processing process,the processing techniques for cluster topology balance and cluster task balance are devised.The devised processing method for cluster topology balance includes consistent hash data distribution and shard number selection for the inverted secondary index clusters.The devised technology for cluster task balance utilizes a weighted directional task allocation strategy which is improved based on diversion polling strategy.The example test results revealed that the topology balance strategy is effective to reduce the time complexity of quasi-real-time data retrieval caused by unreasonable cluster topology planning.Combined with the weighted directional task allocation strategy,the time complexity of concurrent data retrieval in forecasting service is reduced to 100 milliseconds.Furthermore,the probability of distributed deadlock is significantly decreased.Integrating these quasi-real-time data storage management technologies addresses the bottleneck issue due to the conventional data storage technologies.This also enables short-term load forecasting for quick on-demand retrieval of a large amount of quasi-real-time data,hence improves data service level in load forecasting.The study of techniques in the data analysis and mining tier is based on the result of investigating the error components of using supervised learning to forecast load.The stacking heterogeneous ensemble strategy is utilized to combine various heterogeneous forecasting methods to control the changes in population variance which is increased by decreasing the sample dataset deviation.In addition to the conventional zero and primary data-driven framework,the secondary data-driven framework is proposed to reduce the overall deviation of the forecasting model and balance the population variance by deriving and analyzing the related mathematical relations.Furthermore,the example test results also confirm that the stacking heterogeneous ensemble strategy and the secondary data-driven framework improve the quality of the short-term load forecasting model.For instance,the MAPE value of the bus-level load forecasting results of the secondary data-driven stacking ensemble method is as low as 1.736%,which is significantly improved compared with the primary data-driven method.The results also corroborate that the obtained improvements are universal at different situation of data resources.In conclusion,the devised technologies improve the quality of the model used for load forecasting.