Research On Data-driven Optimization Learning Method For Short- And Medium-Term Electricity Demand Prediction

As a form of energy, electric power leads the energy and industrial revolution and changes the people's life from the beginning of the birth. Electric power system is a system which consists of electric power production, electric power transmission, electric power distribution and electric power consumption. It makes the entire power industry joining together effectively and the industry chain is formed then. The accuracy electricity demand prediction is the guarantee of security, stability, economic benefit and high efficiency of power system. The results of electricity demand prediction can not only provide the basis of approving power investment project for the management department, but can also provide decision support for optimal scheduling of power grid and data support for production plan of power stations. The electric energy is different from other consumables, an important feature of electric energy is that it cann't be stored or storage cost is huge. So it requires the electricity be immedietly delivered to the consumers when it is produced and makes production, transmission and consumption of electricity at the same time to maintain the dynamic balance of electric power production and consumption. The precisely electricity demand prediction can adjust electricity production and scheduling plan and avoid the imbalance between production and consumption and provide decision support for decision-maker.Electricity demand data shows different trends in different lengths of time. Most of them are nonlinear, fluctuation and randomness. Thus the traditional forecasting methods can not obtain accurate prediction results. In this paper, different models and methods are proposed to forecast different types of electricity demand based on the power data, the main research work is as follows:(1)For the problem of short-term electricity demand prediction with single factor (Time series), an improved ARMA model with adaptive genetic algorithm is proposed. In this model, the genetic algorithm is used to decide the order of ARMA model. Since the traditional adaptive genetic algorithm has the shortcoming which may cause the stagnation of population evolution when the fitness value is close to the maximum fitness value of population, the adaptive crossover probability and mutation probability are modified and the number of iterations is also taken into the probability. Then the proposed model is adopted to hourly electricity demand forecasting while the data are from PJM INT., L.L.C. The results show that the improved model has better performance on prediction.(2)For the problem of mid-term electricity demand prediction with single factor (Time series), the monthly and yearly prediction of electricity demand are studied. For the monthly electricity demand prediction, a seasonal GATS-SVR model is proposed in the research. In the model, genetic algorithm and tabu search are combined together to search the parameters of SVR model. The hybrid algorithm can overcome the disadvantages of premature convergence and deficiency climbing ability of genetic algorithm. The seasonal adjust index is brought to the model to improve the prediction accuracy. And the least absolute criteria can increase the stability of the model. Then, the seasonal GATS-SVR model is applied to the monthly total electricity consumption of Jiangsu province and the results show that the proposed model has better prediction performance than the compared models. For the yearly electricity demand prediction, a PSO-GM(1,1) model with initial condition optimization based on least absolute criteria is proposed. In the model, the initial condition is improved with the latest data point multiplied by a gain or reduction factor variables compared with original GM(1,1) model. And the least absolute criterion is used to instead of least square criterion for avoiding the error expanding when the singular point exists. The PSO algorithm with time varying weight is used to parameters searching. At the end, the optimized GM(1,1) model is applied to total electricity consumption of China and the results show that the prediction accuracy of this model is superior to the original model.(3)For the problem of mid-term electricity demand prediction with multi factors, the influence of economic and environmental factors associated with the electricity demand are considered and the PSO-CV-SVR model and GA-SVR model are proposed by using the related factors and the historical electricity demand data. In the models, the time-varying weight PSO algorithm and genetic algorithm based on least absolute criteria with cross validation method are used to search the parameters of SVR. The models are used to monthly total electricity consumption of Jiangsu province and the results show that the proposed models are better than BP-neural network model on prediction accuracy.(4) For the problem of mid-term electricity demand interval prediction with multi factors, a WMC-GS-CV-SVR model is proposed. In this model, the GS-CV-SVR model will give the first prediction results with the parameters optimization of grid search and cross validation. Then the weighted markov chain is used to turn the point prediction into interval prediction. The proposed model gives the interval probability distribution for the future data. So that the decision makers can understand the probability information which will be overestimates or underestimate of future electricity demand. Finally, according to the interval probability distribution, the expectation prediction method is used to forecast electricity demand of Jiangsu province and the results show the proposed model has higher prediction accuracy.