Trend Prediction of DC Measuring System Based on LSTM

The accuracy of DC measurement system directly affects the reliable operation of DC control and protection system. In order to improve the estimation and prediction of DC measurement system op-eration state, a trend prediction algorithm based on multi-dimensional analysis and long-term memory network is proposed. Based on the analysis of DC measurement principle, the DC measurement status is diagnosed and abnormal is identified by time series trend analysis and anomaly detection. The LSTM is used to construct a multi factor driving current prediction model, and the model is trained and an-alyzed based on the actual operation data. Compared with the traditional time series prediction model, the results indicate that the proposed method is more accurate, simple and effective, and can be applied to the prediction of driving current.


Introduction
The DC measurement system provides input signals for control and protection which measurement accuracy directly affects the safe and stable operation of the HVDC transmission system in the UHV DC transmission project. At present, the failure of the DC optical measurement system is one of the main difficult problems faced by the operation of the DC transmission system of China Southern Power Grid in recent years. When the driving current or voltage measurement has a large deviation, the protection system may act, which will affect the normal operation of the converter station. [1] The operation status of the DC measurement system is mainly monitored by the personnel's DC measurement data monitoring and threshold judgment methods in the daily operation and maintenance of DC converter stations. It is impossible to predict the operation status of the DC measurement system in advance. And when the DC measurement data does not exceed the alarm value, its abnormality (such as a slow growth trend) may be hidden in the fluctuation of the normal monitoring signal which is difficult to accurately identify it under manual monitoring. Because the DC measurement system has a high acquisition frequency and a large amount of data, simple manual statistical analysis is timeconsuming and labor-intensive. Therefore, it is necessary to carry out forecasting and intelligent analysis for the trend of DC measurement data to provide auxiliary basis for the state evaluation of the DC measurement system.
Multi-feature fusion based on artificial intelligence is an effective means to improve the reliability of big data analysis and defect diagnosis [2], which has been used in the operation and maintenance of converter transformers [3], switches [4], arresters [5] and other equipment. However, there is no corresponding research on the prediction andanalysis of the trend of DC measurement data. Existing research methods for equipment state prediction include time series-based regression analysis models [6]- [7], support vector machine algorithms [8], and deep learning algorithms [9], etc. However, the regression model based on time series is based on linear model, which is difficult to meet the non-linear characteristics of monitoring state quantity of DC measurement system. Although support vector machines have a certain generalization ability, the model prediction accuracy is difficult to improve with the increase of DC measurement characteristic factors. The prediction method based on deep learning enables the model to have non-linear expression capabilities, which can better adapt to the analysis and trend prediction of multi-dimensional data. Compared with the independent characteristics of the calculation results of other neural networks, the calculation results of each hidden layer of LSTM (Long Short-Term Memory) are related to the current input and the previous state, which has stronger adaptability and advantages in the equipment state timing prediction.
In the above context, this paper studies multi-dimensional time series trend analysis and correlation analysis algorithms based on the online monitoring data sources, and applies LSTM to DC measurement data trend prediction. Multi-dimensional analysis of the DC measurement system status evaluation and trend prediction model is established. The defects of the energy transmission fiber and the remote module in the measurement system are identified. Finally, the effectiveness and feasibility of the proposed method are verified combined with the ±800kV DC measurement system to collect field data.

Analysis of DC Measurement Principle
The structure of the DC optical measurement system of the converter station is shown in Figure 1. DC measurement systems generally use photoelectric hybrid methods to collect and transmit signals. The merging unit includes a laser driver and an I/O module. The laser driver supplies power to the remote module through an energy supply fiber, and the I/O module receives data collected by the remote module through the optical fiber.
After the measured state quantity of the primary equipment is stepped down, the analog signal of the state quantity is converted into an optical signal by the remote module, which is transmitted to the merging unit by optical fiber, and the data is transmitted to the control and protection system according to a certain communication protocol.
The causes of abnormal DC measurement data include the abnormality of the measuring device and the abnormality of the DC system itself. The DC measurement data includes drive current, data level and the temperature of the remote module. The overhaul test for the measuring device includes current ratio/voltage ratio check, voltage divider resistance, voltage divider capacitance, laser power, fiber attenuation test, etc., as shown in Table 1. Because periodic test data cannot be obtained online in real time, and it is difficult to meet the needs of online real-time analysis of the system, this paper focusing on analyzing the driving current and data level status based on online monitoring data, and providing decision-making basis for operation and maintenance personnel to inspect.

Multi-dimensional evaluation algorithm
The DC transmission system is generally equipped with AB two-sided screen cabinets under the same pole, and 4 sets of monitoring devices are configured for the same monitoring point. The key state quantities collected include drive current and data level. This paper uses vertical and horizontal statistical analysis methods, time series trend analysis methods, and outlier abnormality identification methods [10] to comprehensively evaluate the status of the DC measurement system. The specific method is as follows.
 Compare the drive current and data level of the high and low ends of the same pole horizontally. When the deviation is greater than the threshold 1 δ , it is judged as abnormal. The 1 δ is a static value, which is the average of the difference between the historical high and low data of the same pole.  The longitudinal analysis of drive current and data level is divided into the following two steps: , the warning. Since this time series has a certain trendbut not obvious periodicity, a dynamic threshold is configured to detect anomalies in the data. Based on the average value, maximum value and minimum value of the data in the past time period T, let } , min{ The box plot method is used to detect the abnormal value in the data batch. When the difference between the abnormal value and the mean value exceeds 3 times the standard deviation value exceeds, the data batch is judged as abnormal.
In addition, the outlier abnormality detection is performed by calculating the local outlier factor (LOF) of the data set and comparing the threshold value based on the density of the data set. The specific flow chart is shown in Figure 2.  Identify abnormal measurement points through outlier detection method, and check whether the corresponding energy transmission fiber and remote module are normal.

Prediction of the trend of drive current based on LSTM
The voltage is relatively stable, and the driving current is dynamically adjusted with the operating power of the systemin the HVDC system. Therefore, this paper focuses on analyzing the trend state of the driving current. LSTM controls the input of information by introducing a gating mechanism, selectively adding new information and forgetting previously accumulated information in terms of state quantity predictions affected by multiple factors, so LSTM has stronger adaptability and generalization ability in dealing with non-linear and temporal prediction problems.
 The principle of LSTM LSTM network model includes forget gate, input gate and output gate. The input gate layer takes the input xt of the current layer and the output ht-1 of the hidden unit at the previous moment as input, and the output result it is the information to be updated. The forget gate layer uses ht-1 and xt to output ƒt.
) ( Then update the cell state Ct, where t C is the candidate value.
The output gate layer runs a sigmoid layer to determine which part of the cell state is output. The update of the hidden statehtconsists of two parts, the first part is the acquisition of ot, and the second part is composed of ot and the activation function.
Wherein, σ and tanhare activation functions. Wxi, Wxf, Wxc, Wxo arethe weight coefficients from the input layer to the hidden layer. Whi, Whf, Whc, Who arethe weight coefficientsof recursive connections.bi, bf, bc, bo are bias factor.  Implementation process This paper fit the relationship between temperature, operating power, data level and drive current to achieve the prediction of drive current based on the LSTM model. The specific training process is shown in Figure 3.

Analysis of calculation examples  Multi-dimensional evaluation
This paper collects the actual DC measurement data of a converter station as a sample set. The bipolar DC transmission system includes 180 measurement points, and the sample set covers the operation and maintenance transcript data from December 2018 to November 2019, the sample interval is 1 hour. The state quantities include drive current, data level, temperature, and active power. The monitoring trend of the DC measurement drive current of the polar 1 is shown in Figure 4.  Fig. 4 Monitoring data of high-side drive current of polar 1 First, analyze the data. The monitoring value of drive current and data level of the same measuring point does not exceed the threshold 65mA and 70mV, and the quantitative monitoring value of trend does not exceed the dynamic threshold range, and the system status is normal. The difference | | x x t − between the monitoring value of drive current and the average value does not exceed 3 times the standard deviation value, the drive current indicator is normal. The LOF judgment threshold is determined to be 1 through the historical abnormal drive current set. When the value of k is greater than 50, the detection accuracy of the data reaches 100%, and the detection error rate is 0%. The entire sequence was tested on this basis, and no outlier driving current data was found.
 Prediction ofdrive current Based on the sample data collection of a converter station, a model and simulation environment based on the LSTM algorithm is built, the data is trained offline, and the trained model is encapsulated and embedded in the early warning system.
In this model, ADAM optimization algorithm is adopted, the initial value of the model learning rate is 0.01, the input dimension is 4, the output node is 1, the error threshold is 2×10 -2 , the number of iterations epoch is 80, dropout is 0.2, batch size is 50.When the training error is less than the error threshold, the output drive current value is obtained, and the training ends.
Take the driving current of ameasuring point as an example for analysis, the graph of MAE of the test set and training set during 80 iterations is shown in Figure 5. After training for 40 epochs, the loss remained stable at 1.83%.   Fig. 6 Forecast results of different models It can be seen both prediction models can simulate the changing trend of the drive current. The LSTM prediction model can track the change trend of the driving current very well, but there is a small deviation in the fitting of the true curve at the peaks and troughs, and the overall prediction error is 2.6089%.There is a large deviation between the predicted result and the actual value in the ARIMA model when the actual value of the drive current fluctuates and continues to decrease. The MAE evaluation index result of the test set of ARIMA is 36mA, which is much higher than the 9.27mA of LSTM.
The ARIMA model relies on the regularity of the data or the data after the difference, ARIMA is also greatly affected by the random fluctuation data, and ARIMA does not have the ability to "selective forgetting", so the accuracy of the ARIMA model under this data set is not high enough. In contrast, LSTM has lower error and higher accuracy, which can better fit the change trend of the actual drive current result value.

Conclusion
A prediction algorithm based on multi-dimensional analysis and LSTM network for the problem of state estimation and trend prediction of DC measurement monitoring quantity is proposed.The following conclusions can be obtained through theoretical research and example analysis.
 The drive current and data level of DC measurement are diagnosed by time-series trend analysis and outlier detection methodsto assist personnel in decision-making and the status maintenance of the energy transmission fiber and the remote module on the basis of the existing operation and maintenance threshold diagnosis.Algorithm application is simple and effective.
 A multi-input-single-output LSTM driving current prediction model is constructed combined with the operating power, which taking the actual DC measurement data of the converter station as a sample.The MAE is reduced by 33 mA compared with the ARIMA model under the same sample set, which verifies the validity and feasibility of the model.