Short-term Power Load Forecasting Based On Similar Days And Extreme Learning Machine

Short-term load forecasting of power system is an important work to realize economic operation of power grid. It is an important part of power system planning, and also be the premise and basic of reliable and economic operation of power system. Through accurate forecasting, the power generation plan can be made scientifically, the power plant can arrange the on-off operation of machine set reasonably, the profits of power system will be improved. Thus, research on the methods of load forecasting to improve the accuracy of it is significant. The dissertation mainly studies load-analysis and forecasting arithmetic.In this dissertation, firstly introduce the characteristic of power load, and deeply analyses the effect of temperature, rainfall and time to short-term load. The dissertation researches the method to find out similar days based on fuzzy clustering, which can optimize the sample of load forecasting. This method choose the influence items of short-term load such as weather factors, week type, date type as the basis of similar days. The fuzzy rules are applied to get the items quantized and establish the mapping table, and similarity coefficient method is used to calculate the similarity between history days and the forecasting day. The similar days are chosen based on the clustering level which takes weather and other items into consideration, it reflects the’periodicity’and’more close, more similar’rules of the similar day chosen. This method can find out the similar days effectively.Secondly, based on reasonably finding out the similar days, the dissertation researches a forecasting model which combines the similar days and the Back Propagation Neural Network. The model chooses similar days by comparing the similarity between history days and the forecasting day. The similar days are used as the training sample of the BP Neural Network, and the self-learning and self-adaptive ability of the BP network avoid the subjectivity of the similar day forecasting. By analyzing the measured data from Jiangsu power grid and comparing the result of similar day BP network with the single BP network, it shows the similar day BP network model is more accurate than single BP network.Finally, aiming to the drawbacks of over-fitting and easily getting stuck into local extremes of BP Neural Network, a model based on similar day, wavelet transform and Extreme Learning Machine is put forward. This model decomposes the time sequences of the similar day load into high-frequency random part and low-frequency basic part. To the low-frequency part, forecast with Extreme Learning Machine, and to the high-frequency part, forecast with average method, the final result is the refactoring of these two parts. The wavelet transform separate the high-frequency random part from the power load sequences, which weakening the randomness of power load and provides a filtering effect. After researching the measured data of power grid, the predicted result of W-ELM model is compared with BP network model and ELM model. The result shows the model based on similar day and wavelet-ELM is more accurate and quick than other models.