Two-stage Time Series Data Prediction Model Based On FCM And LSTM

Time series data refers to a kind of series data that arranges the values observed by the system in chronological order.Time series data widely exist in people's daily life,such as the regional power consumption of the power system,the pollutant emission in the meteorological system,the change of stock index futures in the financial field and so on.This kind of time series data is analyzed and studied,and the prediction model is established to predict the change trend and amount of data in the next period of time.Accurate and reliable prediction results have very important guiding significance for the future decisionmaking and development of the system.Therefore,the research on time series data prediction has a wide range of application scenarios and long-term practical significance.However,most of the existing models can not explain the prediction results,which has a great impact on the subsequent decision-making.As a discrete and nonlinear system driven by data,fuzzy cognitive map(FCM)combines the advantages of fuzzy logic theory and neural network model,and has strong knowledge representation and logical reasoning ability.At the same time,fuzzy cognitive map also has the ability to mine the causal relationship between data variables.It can mine and learn the causal relationship from the past time series data,and then truly reflect it to the objective world through its own weighted directed graph structure.This feature makes FCM can realize the accurate description of dynamic system.It has important application value in prediction and evaluation.Based on the existing research work,the main research contents of this paper are as follows:(1)The establishment method and learning algorithm of fuzzy cognitive map are studied.According to the interpretable and systematic characteristics of fuzzy cognitive map,analyze and predict the time series data and learn the causal relationship between different factors.(2)The prediction ability of different single models to time series data and the analysis and prediction ability of mixed models are studied.(3)The research uses the characteristics of fuzzy cognitive map to establish the causal relationship between data factors,so that the prediction results of the hybrid model can be explained to a certain extent.In summary,the main innovations of this paper are as follows:1.An improved structure optimization construction algorithm(SOGA)is proposed to construct the structure of fuzzy cognitive map,and a new error function is proposed,which will punish the fuzzy cognitive map with high complexity.2.A time series data prediction model connecting FCM and LSTM is proposed.The model takes the concept node in the fuzzy cognitive graph as the input vector of the LSTM network structure.The concept node here is selected by the proposed structure optimization algorithm.In this way,the information connection between the two networks is realized.Based on the above research objectives,in order to verify the effectiveness and feasibility of the research content proposed in this paper,the design is verified by comparative experiments.According to three different time series prediction needs,the corresponding historical data sets are designed,and the methods in this paper are compared with other methods,which are: the data prediction of national total logistics,the data prediction of urban water demand in Jinan,and the daily closing point prediction of Shanghai and Shenzhen 300 stock index.The experimental results show that this method selects the most important concept nodes and edge connections for the target prediction results in the data set,and achieves good data prediction accuracy.After experimental verification,the method proposed is compared with other standard model methods,and the experimental results are analyzed.The results show that the method proposed in this paper has higher prediction accuracy compared with the benchmark method.At the same time,it can reduce the complexity and redundancy of fuzzy cognitive map and enhance the readability and interpretability of the model.Moreover,for large-scale systems with high complexity,the implementation results of this method will be more obvious.