636 / 2019-04-12 10:46:20

Short-term Load Forecasting Method Based on Dual-tree Complex Wavelet Grey Relational Analysis and Deep Self-coding Neural Network

short term load forecasting,Double Tree Complex Wavelet (DTCW),Grey Relational Analysis (GRA),Deep Self-coding Neural Network (DSCN)

The load capacity of the power grid is related closely to the operation of the power system. Accurate and timely power load forecasting plays an important role to the safe operation and the optimal resources allocation and utilization of the power grid. Therefore, power load forecasting is an important aspect of power system research. Because the power load data is essentially a non-stationary random sequence with high complexity and non-linearity, it has been the research focus to improve the prediction accuracy. For short-term load forecasting, the affecting factors include not only the characteristics of load data, but also the external meteorological factors such as temperature, humidity and wind speed. Consequently, it will lead to low generalization ability and inaccurate results by only taking power load data as a series variable to forecast and ignoring the influence of environment. Considering power load data and external environmental factors, a short-term load forecasting method based on Dual-tree Complex Wavelet (DTCW) Grey Relational Analysis(GRA) and Deep Self-coding Neural Network (DSCN) is proposed.
First, the original load data is decomposed into load subsequences by DTCW. Compared with traditional wavelet analysis, DTCW has many excellent characteristics, such as shift invariance, frequency aliasing suppression and approximate analysis. DTCW decomposes load into subsequences, which reduces the complexity and instability of data and overcomes modal aliasing.
Considering the influence of meteorological factors such as temperature, humidity, wind speed, precipitation and air pressure on short-term power load forecasting, grey relational analysis (GRA) is used to extract meteorological factors with strong correlation with load data, which can not only fully characterize the impact of meteorological factors on the load change, but also overcome the high dimension of meteorological data, so it can reduce the running time.
Each subsequence and its strongly correlated meteorological factors are taken as a new feature sequence. According to the contribution ratio of each new sequence, the main load sequence is selected and input into the deep self-coding neural network for short-term load forecasting. Deep self-coding neural network has strong generalization ability and can be applied in different fields. It has good fitting ability for complex data and can improve the accuracy of prediction results.
Experiments were made on the pollutant and meteorological data of Shijiazhuang in Hebei Province, and the electric power data of a certain area in the south of the United States. It is a bright point that, compared with BP neural network prediction and firefly optimized BP neural network, the load prediction based on the model in this paper improves the prediction accuracy more effectively.