Variable Input Structure User Load Forecasting Method Based on User Load State Identification

Along with the roll-out installation and continuous improvement of smart meter data acquisition systems and distribution automation systems in smart grid in recent years, smart meters have been widely installed and used to acquire massive load data in distribution and utilization side, which laid a valuable data base for load forecasting. Due to the significant difference between load in system level and load in distribution and utilization side, traditional load forecasting methods based on the characteristic of system load are typically not appropriate for load in distribution and utilization side, thus requiring load forecasting methods designed in accordance with the special load characteristics in distribution and utilization side. This thesis builds a four-level spatial architecture for load in distribution and utilization side, which is ‘customer-customer cluster-bus-virtual bus’, and proposes a framework for methodologies and applications of load forecasting in distribution and utilization side. Adapted to the massive volume and strong volatility of customer load, this thesis proposes a customer load data compression and load forecasting method based on state analysis. The method mines key load features through load state analysis, then designs a highly efficient data compression format which restores key load features rather than the original load, decreasing the data volume significantly. Based on the compressed data, a variable input support vector machine based on state analysis is proposed to forecast short-term customer load. This model searches and identifies the historical load which is in the same state with the forecasted load as the input of the model, decreasing the adverse effect on load forecasting yielded by customer load randomness and improving the accuracy of short-term customer load forecasting.