A Fuzzy Quadratic Surface SVQR For Probabilistic Electric Load Forecasting

Electricity is an important kind of resources,which has closely relationship with national economy and people's livelihood,and plays a vital role in people's lives and industrial production.With the continuous progress of society,the scale of cities has been gradually expanded,and industrial production has also been developing dramatically.So,the entire society has an increasing demand for electricity resources.It is the duty and obligation of power companies to ensure the timely supply of electricity to the society.In order to get knowledge of the regional electricity consumption for the future period of time,in addition to the fact that electricity resource is hard to be stored,companies often predict the electrical load and use the forecast results to guide their production.Improving the accuracy is power companies'main aim.On the one hand,if the load forecasts are too low,social demand will not be met timely,leading to a blackout accident and causing great economic losses;On the other hand,If the load forecasts are too high,well resources will be wasted,and if the peak is predicted too high,the requirements for power transmission equipment will increase accordingly.However,in fact,the electrical load is always affected by factors such as changing temperature and wind,so the results of point forecasting are more and more difficult to meet the accuracy requirements of power companies.Therefore,the academic and industrial circles are gradually turning their attention to probability forecasting.Probabilistic forecasting can provide an accurate forecast interval,bringing more information to business managers for decision-making,rather than a single forecast value,bringing limited information.The more accurate the forecast,the more valuable it is,the manager can make more correct decision.With the aim of improving the accuracy of power load forecasting,through reviewing the relevant domestic and foreign literatures on power load forecasting,it is found that:1)In the field of power load forecasting,the future temperature is an important influencing factor,so the temperature forecasting should be incorporated into the power load forecasting solution,but there are few relevant studies on it at home and abroad.2)At present,the research on load probability forecasting model mainly focuses on the hybrid model formed by the combination of multiple models,and the research on single probability forecasting model is relatively rare.Therefore,in order to further improve the prediction accuracy and efficiency,this research proposes a new model and solution.The main tasks are as follows:(1)The original kernel-based support vector quantile regression model has been redesigned into a fuzzy quadratic surface support vector quantile regression model.We don't use kernel function to map the data from low-dimension space to high-dimension space,but use a quadratic surface to fit the data directly,which reduces the complexity of the model dramatically.Besides,we designed a new fuzzy membership function in order to assign different weights to different sample points,reducing the influence of noise points on model fitting.(2)Since temperature is the main factor affecting electric load and the future temperature is unknown,it is necessary to predict temperature precisely.We first constructed temperature's feature data-set,and then predicted the probability of temperature at five quantiles respectively.At last,we obtained the temperature prediction interval so we could simulate the possible future temperature scene.(3)Through analyzing electric load data-set,we explored and constructed the feature data for probabilistic electric load forecasting.On this basis,the fuzzy quadratic surface support vector quantile regression model was used for probabilistic electric load forecasting under the nine quantiles and obtained a prediction interval.Compared the model with three other classical probabilistic forecasting models at the aspect of efficiency and accuracy.Finally,it is verified that the new model is prior to the other three classical model.The conclusions of this research are as follows:(1)The proposed and applied fuzzy quadratic surface support vector quantile regression model(Fuzzy QSSVQR)has higher accuracy and efficiency than the other three commonly used probabilistic forecasting models.Instead of using a kernel function to map the data from low-dimension space to high-dimension space,we applied a quadratic surface to fit the data at original space directly,breaking the limit of the kernel,reducing the complexity of the parameters adjustment and improving the model's interpretability.Compared with QR,QRNN and KSVQR,the new model can achieve higher accuracy and efficiency and provide guidance for the production and operation of power enterprises.(2)In temperature forecasting,month,hour and historical temperature all can be used as feature data of temperature for prediction in the future period of time.First of all,the temperature obeys the rule of periodic changes,and it varies greatly between different months;Secondly,it is obvious for the difference of the temperature between different hours of the day.The night temperature is significantly lower than the daytime;Finally,we pictured the relationship between temperature and historical temperature in the same period on the monthly and hourly level,and it was found that there was a very obvious linear correlation between them.Therefore,historical temperature in the same period can also be used as a feature data for temperature prediction.