Research On Distributed Data Warehouse And Data Mining For Load Analysis And Forecasting

Accurate power load forecast plays a significant role in planning and operation of power system, but the load forecast methods applied nowadays can't fully meet the requirements of power system. At present, most of power load forecasting reaserch is focus on improvement in some existing forecast methods or exploring application of new mathematical methods, without considering much of real characteristics of power load, so it is limited in improving forecast accuracy. Now it is rough in thinking of meteorological factors, that is, the same weather condition is generally applied in the wide range of estimation area, which is not enough for large-scale power grids with wide power supply area. In this thesis, according to features of power grid operation and management, mesh partion is made to power grid on basis of region, distributed data warehouse including power load and influence factors is built, and impact of meteorological factors on power load is deeply analysized. Also, the gridding power load analysis and forecast methods based on data mining is reaserched, and the load forecast accuracy is improved through improving power load forcast agorithm and strategy. The main work of the thesis is as the followings.An idea of gridding load and meteorological sensitivity analysis based on distributed data warehouse is prensented, and corresponding analysis software is developed. First, according to natural region, gridding partition is made to large-scale power grid and distributed data warehouse modal with the theme of analysis and forecast is built. Secondly, four comprehensive meteorological indicies are introduced to detailedly analysize relevance and sensitivity of weather and load relevant to each sub-mesh. Then comprehensive meteorological index is worked out by weighted load meteorology sensitivity of every area, which is the base of forecasting next year's load and meteorological index sensitivity variation curve.A diurnal characteristic-load forecast method is proposed, which is based on improved decision tree mining algorithm with cosidering optimization of Property-Value to information gain. In this method, meteorological properties discretizing is made with cluster analysis and information entropy. With this method, the relation between weather and load can be considered more objectively, which can control the amount of discretization break point within a reasonable value; the information gain optimization with considering Property-Value can make up the deficiency of ID3 algorithm, that is, it can reduce the decision tree's depth and improve the query speed and effeciency, then finally achieve the purpose of improving diurnal characteristic-load forecast precision. Also, the test comparison is made in the load forecast example and proves the effectiveness of the proposed method. A gridding short-term load forecast method based on MDRBR (mining default rules based on rough set) is proposed. With this method, on the one hand, the redundant rules introduced by noises can be reduced, which makes classification rule set much smaller and improves the efficiency of rules'production and classification; on the other hand, according to regional features, the meteorological factors and load features of sub-mesh can be considered more detailedly to form different gridding load modal and improve large-scale power grid short-term load forecast accuracy. Then the gridding and non-gridding load forcasting results are compared to show the effectiveness of the method in increasing forecast accuracy which is proposed in the thsis.