Research And Application Of The Time Series Forecasting Models Based On Artificial Intelligence Optimization

Forecasting can reveal the intrinsic contact and development trend of the ob-jective,draw the primary outline of the objective in the future and provide sufficient scientific evidence for decision-making on the basis of the objective rule understanding.As is known,without observation,there is no fitting,while without fitting,there is no forecasting.By using the time series forecasting and fitting forecasting models,this research aims to study the following two aspects:one is the improvement of the time series forecasting model,and the other is the research about the fitting forecasting model with regard to a special time series which is named the probability distribu-tion.Focusing on these two aspects,four time series forecasting and fitting forecasting models are proposed:(1)The seasonal item and the trend item are usually coexisting in the real time series,however,the existing time series forecasting models mainly focus on the trend item but ignore the seasonal item.In this research,an exponential smoothing forecasting model based on the time series decomposition and the particle swarm optimization(PSO)algorithm is proposed:we first decompose the time series into seasonal and trend items by using the time series decomposition method,and then forecast the trend item by using the original first-order and second-order adap-tive coefficient models and the first-order and second-order adaptive coefficient models optimized by PSO.(2)Although neural network models can obtain high forecasting accuracy,the accuracy is not as high as desired when the neural network model is used to directly forecast the time series with the seasonal item.Thus,similar to the first model,this research also develops a back-propagation neural network forecasting model based on the time series decomposition approach.(3)Weibull distribution has always been used to fit the probability distribution of the data,however,the existing traditional parameter estimation methods cannot obtain a high estimation accuracy.Thus,this research uses the PSO algorithm and the differential evolution algorithm to estimate the unknown parameters in the Weibull distribution,and develops a model named the Weibull distribution model based on the artificial intelligence parameter optimization algorithms by constructing different loss functions.(4)In general,the fitting forecasting accuracy evaluation criterion may replace the actual probability by the frequency,or it will make no sense in the case of the denominator equals to ze-ro.To make up these deficiencies,in this research,the continuous ranked probability score(CRPS)is introduced to the probability density fitting forecasting of the time series so as to make up these deficiencies.And on the basis of the definition of CRPS,the analytical expression corresponding to the two-side truncated normal distribution is obtained.The conclusion can be adopted to cases in which there are two arbitrary truncation points,and it is the extension of the conclusion for which the distribution function is non-truncated normal distribution or the one-side truncated normal dis-tribution whereas the left truncation point is zero.When the analytical expression of CRPS is obtained,this research presents five different truncation points estimation approaches on the basis of the three sigma rule as well as the six sigma rule.Finally,the cuckoo search algorithm is used to estimation the unknown location parameter and scale parameter.To verify the effectiveness of these models,we apply them to the following fields:(1)Regarding the exploitation and application of the new energy as the entry point,the first two models are used to forecast wind speed of different sites in the Hexi Corridor of the Gansu province.(2)Since the wind distribution plays an important role in the site selection of the wind farm,the third model is used to the wind speed distribution fitting forecasting of the Inner Mongolia.(3)Anomaly detection and disposal is an important task in data mining and data processing.Thus,this research uses the fourth model to the anomaly detection of two actual data series:Iris data series and fourclass data series.All of these forecasting and fitting forecasting practical application cases demonstrate that the new models proposed in this research can obtain higher accuracy results as compared to the corresponding original models and some other classical forecasting and fitting forecasting models.