Application Of Machine Learning In Load Forecasting Of Mountain Power Grid

Power load forecasting is to predict the development trend of future load through historical data.High accuracy load forecasting can provide decision-making reference for power system planning,operation and dispatching,so as to ensure power supply reliability and improve economic benefits.The following problems exist in the load forecasting of the mountain area: 1)the collection of load data in mountainous areas has low degree of automation,old equipment,difficult communication,and it is easy to lack data or produce some abnormal data;2)Mountain power grid connected to a higher proportion of distributed generation,its load is easily affected by a variety of characteristic factors,directly according to the historical load data prediction accuracy is low;3)Mountain power grid is more vulnerable to the impact of various natural disasters,and its network structure is weak,old equipment,it is likely to have a sudden reduction of load,resulting in its load forecasting historical data is not reliable.In view of the above problems,this paper proposes a data processing method and a load forecasting method suitable for mountain power grid.In the aspect of data processing and effect evaluation,this paper first introduces the data preprocessing methods in load forecasting,and analyzes their shortcomings.Then,the data processing methods suitable for load forecasting in mountainous areas are proposed,including 1)quartile range DBSCAN abnormal data detection method based on particle swarm optimization.2)Two stage missing data filling method based on back propagation neural network expectation maximization.They solve the problem that it is difficult to determine the threshold and initial value in conventional methods.Finally,the importance of introducing the degree of freedom penalty into the evaluation index of prediction effect is discussed.In the aspect of load forecasting of high permeability new energy grid in mountainous area,a short-term load forecasting method based on bidirectional long-term and short-term memory network considering feature selection is proposed for high permeability new energy grid in mountainous area with large amount of feature data.Firstly,the feature data is clustered according to the density,and then the sample data is mapped into the weight induced space.A kind of interval is defined to select the feature and delete the irrelevant feature.Finally,the bi-directional long-term and short-term memory network is used to predict the load of the selected data.Taking a mountain power grid in China as an example,the effectiveness of the method is verified,and the results show that it has better accuracy and applicability than the traditional method.In the aspect of load forecasting of disaster stricken power grid in mountainous area,aiming at the problem that historical data is difficult to directly refer to in load forecasting of disaster stricken and recovery process of power grid,a transfer learning method based on attention mechanism is proposed for load forecasting.Firstly,the historical data is decomposed by empirical mode decomposition(EMD)to obtain different frequency components;then,the learning model of attention mechanism encoder / decoder is established,and the learnable parameters are trained on the load component data set;then,the learnable parameters are transferred to the attention mechanism cif-lstm model,which is used to predict the basic trend components and periodic high-frequency components respectively;finally,the learning parameters are transferred to the attention mechanism cif-lstm model Then,the load forecasting results are superimposed to get the final load forecasting value.Taking the historical load data of Dabie mountain disaster area power grid as an example,the method can effectively solve the problem of insufficient data and improve the prediction accuracy.