Power communication digital flow prediction method based on VMD-LSTM-SVM model

Under the current trend of abundant information on power business, large data concentration, and large flow explosion, aiming at the randomness, volatility, and uncertainty of massive flow of electric power communication network, a digital power flow prediction method based on VMD-LSTM-SVM model is proposed. The interaction between the values of each traffic index before and after time is considered. LSTM is used to process traffic data and make an accurate prediction of future traffic. The power communication network can make dispatch responses to possible communication congestion by using link resources according to traffic prediction results and ensuring the transmission quality of power service data.


Introduction
With the increasing business of smart power grid, the data acquisition network of intelligent distribution and communication network will face great pressure of data flow at the application layer, which will bring network congestion, data transmission rate reduction, network equipment faults increase, and other problems to the power information network.Therefore, it is very important to predict the network traffic of the smart grid accurately and quickly.
The existing research mostly combines neural network models and machine learning algorithms to predict network traffic.Early traffic prediction models, such as support vector machine, Markov chain model, etc., although achieved good prediction effect, cannot dig deeply into the internal relationship between data.Li [1] proposed a diffused convolutional recursive neural network for data-driven traffic prediction, which captures spatial correlation through the deep learning framework of traffic prediction.However, RNNS have the problem of not being able to handle long-term dependencies.Long Short-Term Memory (LSTM) has strong adaptability and excellent scalability for big data training.In reference [2], an LSTM model based on Bayesian optimization is proposed to predict the high-speed traffic flow given the difficulty of manual empirical parameter adjustment.However, the existing LSTM neural networks have high operational complexity and long running time.The traditional traffic prediction model also has the problem of multi-mode aliasing interference between traffic sequences [3] and cannot accurately predict traffic services with long time scales.
Given the above problems, this paper proposes a power digital power flow prediction method based on the VMD-LSTM-SVM model and simulates the historical data set of power communication traffic sequences on the cloud platform to verify the prediction accuracy of the model.It can effectively solve the problem of massive traffic congestion in the power communication network.

Power data service analysis
Data services with the same transmission characteristics are classified into the following three types based on the transmission characteristics of traffic time ductility, transmission rate, link reliability, packet loss rate requirements, and jitter amplitude.
(1) Automatic control business: including SCADA and MIS data business.SCADA requires high reliability and real-time performance, usually with low traffic (300Kbps to 800Kbps) and high delay requirements.MIS data has very high burst traffic (the maximum peak value can reach 4-6 Mbps), while the transmission of the network requires higher bandwidth and does not require a too high delay.
(2) Power dispatching program control service: mainly includes dispatching telephone, administrative office telephone, and conference telephone.Voice services require high reliability and real-time performance, and do not require high bandwidth.
(3) Remote viewing services: Remote viewing services mainly include video conferences, unattended substations, and other video surveillance, which have high requirements on network time extension and bandwidth.

model framework
For massive power data services, this paper proposes the VMD-LSTM-SVM model.The specific implementation process is as follows: Step 1: The historical data of the cloud platform power communication service traffic sequence is preprocessed by the variational mode decomposition (VMD) method, and the flow sequence is extracted into several intrinsic mode components and residual components.
Step 2: The decomposed intrinsic modal component IMF and residual component R were correlated with different traffic service types respectively, and the Pierson correlation coefficient was used to study the impact of different traffic service types on the accuracy of the prediction model.
Step 3: For the intrinsic mode component IMF after VMD decomposition, LSTM is introduced.In the case of the minimum loss function, BPTT is adopted as the backpropagation algorithm based on time series to determine the optimal parameters of the model, the LSTM sub-prediction model is established, and d k of each component model prediction is output, where k=1,2,3...m-1; Step 4: For the residual component R, combined with the influencing factors of the flow sequence, a support vector machine (SVM) was used for model regression fitting, the optimal parameters were determined, and the predicted value dr was output.
Step 5: After the IMF component and residual prediction models are completed, the predicted values of each model are superimposed and output.Finally, the predicted value d of power communication network traffic is output.

VMD Preprocessing
Variational mode decomposition (VMD) is a new adaptive mode variational extraction method for nonstationary signals, which can realize effective separation of intrinsic mode component (VIMF) and frequency domain division of signals, thus effectively reducing the non-stationarity of complex nonlinear time series [4] .Therefore, the VMD method is adopted in this paper to preprocess the historical data of the power communication service traffic sequence of the cloud platform, and the traffic sequence is extracted into multiple intrinsic mode components and residual components, to reduce the difficulty of subsequent LSTM neural network training.
The core idea of VMD is to construct and solve variational problems [5][6] , assuming that the initial signal is composed of m mode, each modal eigenfrequency corresponds to a   and a finite bandwidth, each pattern represents a frequency modulation amplitude modulation signal   .We initialize all modal number m and eigen frequency   , VMD decomposition by traffic bandwidth constraints, and the corresponding constraint variational model expression [5] : Where, f(t) represents initial signal,   is the intrinsic frequency   corresponding to the first k mode, and   is the real value of   .
To minimize the objective function, the enhanced  Lagrange operator is introduced uniformly constrained into the objective function by type (2): Where  is a penalty factor to ensure the accuracy of the reconstructed signal, and   is the Lagrange multiplier with time.
We calculate the L minimum value, the use of the alternating direction multiplier method is optimized, and the   and   are updates.Alternately, the solution domain is mapped to frequency by the Fourier isometric transform, and the number of iterations for the n: Where ˆ( ) f  represents the initial signal value in the frequency domain, and ˆ( )   represents the Lagrange multiplier in the frequency domain.
The optimal solution, namely the central frequency of the IMF component, was finally obtained through iteration, and the original flow sequence was decomposed into m sub-sequences, including M-1 intrinsic mode function IMFs and a residual component: Where   and r(t) flow sequence of eigenmodes are influenced by history traffic sequence, cyclical change trend, and the trend of the residual sequence is decomposed.
To further improve the prediction accuracy of the model, correlation studies were carried out between the extracted intrinsic modal component IMF and residual component R and different business types respectively, and the Pierson correlation coefficient was used to measure the impact of various businesses on the prediction model accuracy: Where  represents the flow rate of different power data service units in seconds (M/s), and  represents the average flow rate of power data service units in seconds within the sampling period.The correlation analysis results of each component of the traffic sequence and traffic of different data services are shown in Table 1.

LSTM prediction model
LSTM is a special cyclic neural network proposed on the basis of RNN [7] .Compared with traditional cyclic neural networks, LSTM can well solve the problem of gradient vanishing and gradient explosion [8] .
The power business flow sequence m-1 intrinsic mode component IMF after VMD decomposition and the LSTM prediction model are established.
Based on the traditional neural network model, simple neurons in the hidden layer are transformed into LSTM units with the gated mechanism for long and short-duration memory.The unit state of the long and short-duration memory unit is transmitted over time, and the information of the united state is deleted or added through the use of a forgetting gate, input gate, and output gate.The intrinsic modal components IMF1, IMF2, and IMF3 obtained after the decomposition of power digital power flow sequence are taken as input sequences x 1 , x 2, and x 3 , and the output sequences h 1 , h 2 , and h 3 respectively correspond to the predicted flow values d 1 , d 2 and d 3 obtained after LSTM training are input values representing the output values of forgetting gate, input gate, and output gate.In the case of the minimum loss function, the BPTT back-propagation algorithm based on time series is adopted to determine the optimal parameters of the model.

SVM Model Regression Fitting
Support Vector Machine (SVM) is a machine learning algorithm with strong generalization ability and outstanding advantages in solving problems such as small samples and nonlinear and high-dimensional pattern recognition [9] .Therefore, this article builds a residual prediction model of residual error component R using SVM model regression fitting, determining the optimal parameters and output forecast  .
When solving regression problems such as data prediction, the training set sample is given, where x n m i R   represents the input mode and n i y R  represents the target output.First, map the residual component decomposed by the VMD variational mode to the high-dimensional feature space Ψ via the nonlinear mapping function   .The optimal regression plane Ψ has the smallest distance to the plane.The optimal regression plane function is set as follows: Where w T represents the weight vector and b is the bias vector.
According to the criterion of minimum bias distance of regression surface on training data, a constrained function optimization problem was derived [10] : Where ξ is the relaxation variable, ξ* indicates that the standard relaxation quantity is the reflection of the deviation degree of the ideal case, ε is the insensitive loss function, C is the regularization parameter, which is used to punish large migration samples, achieve the balance between large migration sample proportion and model complexity, and strengthen the prediction ability of the model to unknown flow, Lagrange multiplier method.The solution of the optimal regression surface is transformed into an optimization problem with constraints: . 0 0 , 1,..., Where,   1 n i i   is the Lagrange multiplier, and a kernel function is       According to different power business flow data types of Pierson correlation coefficient is set to solve based on the SVM regression fitting, the optimal parameter [10] , the output residual component forecast  .

Simulation and algorithm performance analysis
In order to verify the effectiveness of the VMD-LSTM-SVM model in the prediction of electric power digital power flow, this paper uses MATLAB R2019 as a simulation tool to predict and evaluate the historical traffic sequence of 10kV and below urban distribution communication network in an urban district of Tianjin.We set the learning rate of the LSTM model as 0.001, the number of hidden layers as 2, the number of nodes of each hidden layer as 100, and the IMF number of intrinsic mode components as 3. RNN, LSTM, and VMD-LSTM models were used as comparative experiments, and three indexes of mean absolute error (MAE), mean relative error (MAPE), and root mean square relative error (MSE) was used to evaluate model performance.The traffic prediction results were obtained as shown in Figure 1.  Figure 2 shows that, compared with other methods, the predicted value of this method in the flow prediction trend has a higher approximation to the actual value.In the automatic data service, the mean absolute error (MAE) is reduced by 2.02%, in the power dispatching service, the root means square relative error (MSE) is reduced by 1.67%, and in the remote viewing service, the RMS relative error (MAPE) was reduced by 0.12%.
At the same time, SVM regression fitting enables the model to carry out regression analysis on the residual components of different types of digital power flow, which further improves the accuracy of the model.

Conclusion
Aiming at the problem of network congestion caused by massive power business data traffic, this paper proposes a power digital power flow prediction method based on the VMD-LSTM-SVM model.Through VMD decomposition and preprocessing of the traffic sequence, the multi-mode aliasing interference of the traffic sequence is solved, the running complexity of the LSTM neural network is reduced, the running time is shortened, and the prediction error is reduced.At the same time, based on the correlation analysis results of each flow sequence for different digital power flows, the residual component of power business flow was analyzed by SVM regression, which effectively reduced the prediction deviation caused by the data flow in the burst current and further improved the accuracy of the model.

Figure 1 .
Figure 1.Comparison of experimental results of traffic prediction by different algorithms The data flow errors of the three types of power business are compared respectively for different models, and the simulation results are shown in Figure 2.

Figure 2 .
Figure 2. Comparison of power service data flow errors of different models