A prediction method for the dynamic thermal capacity limit of transmission lines that can track the time-varying characteristics of environmental factors

Abstract

The present invention provides a method for predicting the dynamic thermal capacity limit of a transmission line that can track the time-varying characteristics of environmental factors, and belongs to the field of transmission network line operation status assessment in power systems. First, the main influencing factors of DTR are determined; and a DTR prediction model is established that takes into account the time-varying characteristics of the degree of influence of environmental factors. Secondly, the DTR time series prediction model training data is obtained and the DTR time series prediction model is trained. Finally, the DTR time series prediction model is used to predict DTR, and finally a rolling prediction is achieved. The ForecastNet prediction model with a time-varying weight structure adopted by the present invention can effectively overcome the problem of time-varying environmental input feature weights in traditional time series prediction models, dynamically track changes in the degree of influence of environmental factors, and improve the prediction accuracy of DTR.

Description

Technical Field

[0001] The present invention belongs to the field of transmission network line operating status assessment in power systems. It uses historical big data modeling to predict the dynamic thermal capacity limit of transmission lines over a period of time in the future, and involves a time series prediction method based on the ForecastNet time-varying neural network. Background Art

[0002] Using thermal capacity limits to assess the operating status of power system lines is a fundamental method for evaluating line transmission capacity. Typically, the transmission capacity limit values used by dispatch centers for transmission capacity assessments are static thermal ratings (STR). STRs are thermal limits of line transmission capacity based on extreme environmental conditions and are therefore overly conservative. Dynamic thermal ratings (DTRs), calculated or predicted using actual meteorological information from line operation, more accurately reflect the actual transmission capacity of a line. Therefore, DTR-based dynamic capacity expansion technology for transmission lines can achieve a reasonable expansion of the line's power flow transmission limit within the existing transmission framework, fully tapping the transmission network's potential at a very low cost and improving renewable energy transmission capabilities.

[0003] The thermal capacity limit of transmission lines is closely related to the meteorological environment. According to IEEE Std. 738, the four primary meteorological factors that determine the thermal capacity limit of transmission lines are ambient temperature, wind speed, wind direction, and sunlight intensity. Modern power systems utilize a large number of electrical and meteorological measurement devices that provide robust data support for DTR prediction. Therefore, the forecast values of these four meteorological factors can be used as input features to predict DTR.

[0004] The four meteorological input features used for DTR prediction have the following characteristics:

[0005] 1) All four meteorological input characteristics are time-varying, and some of them show periodic time series changes;

[0006] 2) The four meteorological input characteristics have different impacts on DTR;

[0007] 3) The same meteorological input characteristics have different impacts on DTR at different temporal and spatial scales.

[0008] The existing DTR time series prediction model extracts the time series correlation information of DTR data and the mapping correlation information between environmental data and DTR through cyclic training of historical data. However, the existing DTR time series prediction models are all time-invariant models. The essential reason is that the cyclic prediction unit of this type of model adopts a weight sharing structure. Although the weight sharing structure can overcome the defect of gradient diffusion of the time series prediction model, for DTR prediction, the meteorological environment input that affects DTR is time-varying, and the degree of influence of the meteorological environment on DTR is also time-varying. Therefore, the weight sharing structure assigns the same weight to the same meteorological environment input in different time periods, that is, it assumes that the same meteorological input has the same degree of influence in different time periods, which is unreasonable. Summary of the Invention

[0009] To address the shortcomings of existing DTR time-series prediction methods, this paper proposes a method for predicting the dynamic thermal capacity limit of transmission lines that tracks the time-varying characteristics of environmental factors. This model, based on the ForecastNet neural network, leverages the neural network's time-varying series structure to dynamically track the impact of environmental factors and modify environmental input weights, improving prediction accuracy.

[0010] In order to achieve the above object, the technical solution of the present invention is:

[0011] A method for predicting the dynamic thermal capacity limit of a transmission line capable of tracking the time-varying characteristics of environmental factors comprises the following steps:

[0012] Step 1. Identify the main factors affecting DTR

[0013] The thermal current limit of overhead lines is closely related to the meteorological environment. According to IEEE Std.738, the calculation formula for the thermal current limit of overhead lines is as follows:

[0014]

[0015] Where, I DTR is the dynamic thermal capacity limit of the circuit; T max is the maximum conductor temperature that the transmission line can withstand; q c (T max ) is the air convection heat dissipation of the transmission conductor at the limit conductor temperature, q c It is related to the wind speed, wind direction, ambient temperature and conductor temperature difference in the environment where the line is located; q r (T max ) is the heat radiation dissipation of the transmission conductor at the limit conductor temperature, q r Related to the difference between the ambient temperature and the conductor temperature in the environment where the line is located; q s is the amount of solar radiation absorbed by the transmission conductor at the limiting conductor temperature, q sRelated to the light intensity in the environment where the line is located. R(T max ) is the resistance of the transmission conductor at the limiting conductor temperature.

[0016] The above theoretical calculation formula shows that the environmental factors affecting the DTR of transmission lines include ambient temperature, wind speed, wind direction, and light intensity. Therefore, in order to predict DTR, it is necessary to know the predicted values of the ambient temperature, wind speed, wind direction, and light intensity of the DTR transmission line in the future. It is worth noting that different meteorological factors have different degrees of influence on DTR. On a shorter time scale, wind speed has the greatest impact on DTR, while light intensity has the least impact. Furthermore, the same environmental factor has different effects on DTR at different times of the day, and the degree of influence of various meteorological factors on DTR varies over time. To ensure the accuracy of DTR prediction, it is necessary to account for the temporal and spatial heterogeneity of input features in the prediction.

[0017] Step 2. Establish a DTR prediction model that considers the time-varying characteristics of the impact of environmental factors

[0018] LSTM improves the memory ability of the cycle unit to historical data information through input gate, output gate and forget gate. Its cycle unit structure is as follows Figure 1 As shown, the mathematical model is as follows:

[0019] m t =σ([W m ·h t-1 ,V m ·i t ]+b m ) (2)

[0020] d t =σ([W d ·h t-1 ,V d ·i t ]+b d ) (3)

[0021]

[0022]

[0023] u t =σ([W u ·h t-1 ,V u ·i t ]+b u ) (6)

[0024] h t =u t ⊙tanh(c t ) (7)

[0025] Where h t is the hidden layer output of the LSTM recurrent unit at time t; σ is the activation function; ⊙ represents the bitwise multiplication of the elements at corresponding positions of the two matrices; b is the bias of each gate calculation unit. Define W0 and V0 as the hidden layer output data h of the previous period respectively. t-1 The input data of environmental meteorological characteristics during this period i t The weight coefficient matrix of W0 and V0 is shown in formula (8) and formula (9) respectively:

[0026] W0=[W m ,W d ,W c ,W u ] (8)

[0027] V0=[V m ,V d ,V c ,V u ] (9)

[0028] In the formula, the elements represent the weight coefficients of the forget gate, input gate, hidden layer state and output gate in order.

[0029] According to the above formula, the mapping equation of the cyclic unit of the classic DTR time series prediction model can be simplified as follows:

[0030] h t =f0(W0·h t-1 ,V0·i t ) (10)

[0031] Where f0 is the mapping function of the cyclic unit, which is time-invariant in time series prediction; W0 and V0 are also time-invariant weight parameter matrices.

[0032] The above mapping equations show that the classic DTR time series prediction model is time-invariant. This is fundamentally due to the fact that the cyclic prediction unit of this model adopts a weight-sharing structure, which overcomes the gradient diffusion flaw of traditional time series prediction models. However, for DTR prediction, the meteorological inputs that affect DTR are time-varying, and the degree of influence of the meteorological environment on DTR is also time-varying. Therefore, the weight-sharing structure assigns the same weight to the same meteorological input in different time periods, assuming that the same meteorological input has the same impact in different time periods, which is unreasonable.

[0033] Aiming at the defects of traditional models, the present invention establishes a DTR time series prediction model based on ForecastNet neural network. The DTR time series prediction model based on ForecastNet neural network adopts a time-varying weight structure, which overcomes the defect of time-varying weight parameters of traditional DTR time series prediction model. Taking the 24 hours before the day as the time scale of DTR prediction, the improved ForecastNet prediction model structure is constructed as follows: Figure 2 As shown in the figure, i t The environmental input feature matrix at time t contains the predicted values of the four environmental factors at time t; t is the predicted DTR value at time t.

[0034] Compared with the traditional DTR time series prediction model, the DTR prediction model based on ForecastNet in this invention has the following improvements:

[0035] 1) Series structure, parallel in and parallel out

[0036] The ForecastNet neural network is a multi-step feedforward time series neural network. Instead of using cyclic prediction units, this time series prediction network connects 24 prediction units in series, representing 24 time periods. During model training, the weights and parameters of the 24 series units are derived through training on historical data. During the forecasting process, weather forecast data for the next 24 hours is fed into the forecast model in parallel using a parallel input and output method to generate the DTR forecast for the next 24 hours.

[0037] 2) Weight Time-Varying Network

[0038] The hidden layer of the ForecastNet neural network adopts a time-varying structure, and its hidden layer mapping equation is shown in formula (11):

[0039] h t =f t (W t ·h t-1 ,V t ·i t ,U t o t-1 ) (11)

[0040] Where, f t is the mapping function of the hidden layer prediction unit, which can be composed of a BP neural network or a convolutional neural network; W t 、V t and U t are the weight matrices of the previous hidden layer state, meteorological environment input characteristics, and the previous level output; h t-1 is the hidden layer state at time t-1; i tInput data for environmental characteristics at time t; t-1 It is the DTR value output at time t-1.

[0041] V t is the weight parameter of the meteorological environment data input feature, which is obtained by training historical data. It can realize the dynamic tracking of the impact of the input feature. The meteorological environment input feature has different impacts on DTR at different times of the day, which causes V t The value is different in different time periods. t and U t It is also a time variable that can reflect the time correlation characteristics of DTR.

[0042] Through the above-mentioned DTR time series prediction model based on ForecastNet neural network, a DTR time series prediction model that can track the time-varying characteristics of environmental factors can be established.

[0043] Step 3. Obtain DTR time series prediction model training data

[0044] The training data for the DTR time series prediction model consists of input data and output data, both of which are sourced from a historical database. The input data is obtained from a historical numerical weather prediction (NWP) database, where ambient temperature, wind speed, wind direction, and light intensity for each time period are read from the NWP database to form the independent variables of the training set. The output data is obtained from a historical power grid monitoring database, where historical DTR values are calculated using real-world weather data and transmission line measurement data to form the dependent variables of the training set.

[0045] Step 4. Train the DTR time series prediction model

[0046] The DTR time series forecasting model is trained using training data. First, the neural network type for the hidden layer in the ForecastNet series forecast unit is selected; second, a hyperparameter optimization method is selected; finally, the forecast time scale and forecast time interval are determined, and the input and output data for the corresponding time period are fed into the ForecastNet forecasting model for training.

[0047] Step 5. Use the DTR timing prediction model to predict DTR

[0048] Through model training, a DTR prediction model for a specific route can be obtained. Environmental forecast data for the route's area is obtained from the National Weather Prediction Workshop (NWP): ambient temperature, wind speed, wind direction, and sunshine intensity. The data selection time scale and interval must be consistent with the training model. The feature sequence of the meteorological environment for the next period is input into the model to obtain DTR predictions for each period in the future.

[0049] Step 6. Implement rolling forecast

[0050] In order to realize the rolling prediction of DTR, the present invention integrates the NWP prediction data of the existing period and the DTR measured data into the historical database and uses them as training data to participate in model training, thus preparing for the establishment of the DTR prediction model for the next period.

[0051] Compared with the existing DTR timing prediction method, the present invention has the following advantages:

[0052] The ForecastNet prediction model using a time-varying weight structure in the present invention can effectively overcome the problem of time-varying weights of environmental input features in traditional time series prediction models, dynamically track changes in the degree of influence of environmental factors, and improve the prediction accuracy of DTR. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is the recurrent unit structure of the traditional LSTM time series prediction model.

[0054] Figure 2 It is a ForecastNet time series forecasting model with a time-varying weight structure.

[0055] Figure 3 The MAPE values of the 30-day rolling DTR forecast for 46 transmission lines;

[0056] Figure 4 Comparison of DTR day-ahead prediction results of LSTM and ForecastNet time series forecasting models on a random day in August 2021. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of this application clearer, this application is further described in detail in combination with the following actual cases.

[0058] The case study system is based on the IEEE-39 node system topology. It assumes that all lines in the IEEE-39 system are located in four different weather zones, and the lines are divided into four zones according to the four-quadrant topology. Within the Liaoning Province 220kV transmission network database, four transmission zones were selected based on different weather conditions. An equal number of transmission lines were selected from each zone, and their operating data was used as the IEEE-39 node system transmission line operating data. This operating data includes line electrical parameters, line DTR values, ambient meteorological parameters, and NWP values.

[0059] The first step is to obtain the transmission line historical database.

[0060] The database contains historical NWP environmental forecast data, historical environmental measurement data, and operating data for each transmission line. The DTR data for each line at each historical time period is obtained from the historical environmental measurement data and operating data.

[0061] The second step is to determine the type of cascade hidden layer and the method of hyperparameter optimization.

[0062] The ForecastNet model uses a BP neural network as the cascade hidden layer, and is tuned using the control variable method to minimize the loss function. The simulation system hardware environment is based on a Tesla V100-SXM2 GPU server and is implemented in the Python programming language based on the deep learning PyTorch framework.

[0063] The third step is to determine the training data set and test data set, and perform rolling training and prediction on the ForecastNet time series forecasting model.

[0064] The original training dataset comes from the Liaoning Power Grid's 2020-2021 NWP and DTR database. The training set uses data from January 1, 2020, to August 1, 2021, and the rolling test set uses data from August 1, 2020, to August 30, 2021. Rolling forecasts are performed on a daily time scale with an hourly interval.

[0065] The fourth step is to compare the prediction results.

[0066] To verify the accuracy prediction advantage of the ForecastNet prediction model, the LSTM time series prediction model in the classic DTR prediction was selected for comparison, and two methods were used to predict the DTR for the next 24 hours.

[0067] The mean absolute percentage error (MAPE) is used as the evaluation index to compare the prediction accuracy of the two prediction methods. The calculation method of MAPE is shown in formula (12).

[0068]

[0069] Where, is the true value of DTR; is the predicted value of DTR; n is the predicted number of DTR.

[0070] The DTR prediction of 46 transmission lines in the IEEE-39 node experimental system was carried out, and the MAPE of each transmission line on August 30, 2021 was calculated under the two prediction methods. The error statistics are as follows: Figure 3 shown. Figure 3The following table shows the MAPE values for the 30-day rolling DTR forecast for 46 transmission lines. The statistical results show that the average prediction error of the LSTM forecasting model is higher than that of the ForecastNet forecasting model. The LSTM forecasting error is relatively large on all lines, with Line 31 having the highest average prediction error of 9.336%. Compared to the LSTM model, the ForecastNet model significantly reduces the DTR prediction error on all transmission lines, with the highest average prediction error occurring on Line 13 at 4.28%. In this experimental network, the ForecastNet model achieved an average DTR prediction error of less than 5% for all lines. This is because the ForecastNet model can dynamically track the time-varying characteristics of influencing factors and their degree of influence. Compared to classic time series forecasting models, it has higher prediction accuracy and can provide more accurate and reliable capacity margin for transmission lines.

[0071] In order to more clearly compare the prediction accuracy of the classic prediction model and the ForecastNet prediction model for the DTR in the next 24 hours, this paper selects a transmission line with an STR value of 825A to predict its DTR on a random day in August 2021. After optimizing the control variable method, the hyperparameters of the LSTM prediction model for this line are selected as follows: the number of hidden layer neurons is 100, the number of hidden layers is 3, the dropout learning rate is 0.001, the maximum number of training iterations is 100, and the number of batches is 16; the hyperparameters of the ForecastNet prediction model are: the number of hidden layer neurons is 50, the dropout learning rate is 0.001, the maximum number of training iterations is 150, and the number of batches is 16. The results of the DTR prediction using the above parameters are compared. Figure 4 shown.

[0072] Depend on Figure 4 It can be seen that for this transmission line, the day-ahead DTR prediction results of the ForecastNet prediction model are closer to the actual DTR value than those of the classic LSTM prediction model. This is because the weight sharing structure of the LSTM ignores the different impacts of the meteorological environment during the 24 hours of the day, causing the predicted values of some forecast periods to deviate from the actual values. The ForecastNet prediction model uses a time-varying weight structure to make the predicted values fit the actual values more closely. At the same time, Figure 4 It can be seen that the prediction deviation of the ForecastNet model increases in the last few forecast periods of the day, but the prediction trend is still close to the true value. The prediction deviation in this period is caused by the NWP prediction deviation. The DTR prediction deviation in this part can be reduced by improving the NWP prediction accuracy or shortening the DTR prediction time.

[0073] The simulation results verify the accuracy and effectiveness of the prediction method proposed in this invention.

[0074] The above-described embodiments merely express the implementation methods of the present invention, but should not be understood as limiting the scope of the patent of the present invention. It should be pointed out that for those skilled in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for predicting the dynamic thermal capacity limit of a transmission line that can track the time-varying characteristics of environmental factors, characterized in that: The following steps are involved: Step 1. Identify the main factors affecting DTR; Step 2. Establish a DTR time series prediction model that takes into account the time-varying characteristics of the impact of environmental factors; A DTR time series prediction model is established based on the ForecastNet neural network. The DTR time series prediction model based on the ForecastNet neural network adopts a time-varying weight structure, takes the 24 hours before the previous day as the time scale for DTR prediction, and constructs an improved ForecastNet prediction model structure. The DTR time series prediction model based on ForecastNet has been improved as follows: 1) Series structure, parallel input and output The ForecastNet neural network is a multi-step feedforward time series neural network. This time series forecasting network connects 24 forecast units in series, representing 24 time periods. During model training, the weights and parameters of the 24 series units are obtained through training on historical data. During the forecasting process, the weather forecast data for the next 24 hours is input into the forecasting model in parallel using a parallel input and parallel output method to obtain the DTR forecast value for the next 24 hours. 2) Weight Time-Varying Network The hidden layer of the ForecastNet neural network adopts a time-varying structure, and its hidden layer mapping equation is shown in formula (11): h t =f t (W t ·h t-1 ,V t ·i t ,U t ·o t-1 ) (11) Where, f t is the mapping function of the hidden layer prediction unit, which can be composed of a BP neural network or a convolutional neural network; W t 、V t and U t are the weight matrices of the previous hidden layer state, meteorological environment input characteristics, and the previous level output; h t-1 is the hidden layer state at time t-1; i t Input data for environmental characteristics at time t; t-1 The DTR value output at time t-1; V t is the weight parameter of the meteorological environment data input feature, which is obtained by training historical data to achieve dynamic tracking of the impact of the input feature. The meteorological environment input feature has different impacts on DTR at different times of the day, resulting in V t The value is different in different time periods; W t and U t It is also a time variable, reflecting the time correlation characteristics of DTR; Step 3. Obtain DTR time series prediction model training data; Step 4. Train the DTR time series prediction model; Step 5. Use the DTR time series prediction model to predict DTR; Step 6. Implement rolling forecast.

2. The method for predicting the dynamic thermal capacity limit of a transmission line capable of tracking the time-varying characteristics of environmental factors according to claim 1, characterized in that: The specific implementation process of steps 1, 3 to 6 is as follows: Step 1. Identify the main factors affecting DTR In order to predict DTR, it is necessary to determine the predicted values of ambient temperature, wind speed, wind direction, and light intensity of the DTR transmission line in the future period; and to ensure the accuracy of DTR prediction, it is necessary to consider the temporal and spatial differences of the input features in the prediction. Step 3. Obtain DTR time series prediction model training data The training data for the DTR time series prediction model includes input data and output data, both of which come from a historical database. The input data is obtained from the Numerical Weather Forecast (NWP) historical database, where the ambient temperature, wind speed, wind direction, and light intensity for each time period are read to form the independent variables of the training set. The output data is obtained from the power grid historical monitoring database, where historical DTR values are calculated using real-world weather data and transmission line measurement data to form the dependent variables of the training set. Step 4. Train the DTR time series prediction model The DTR time series forecasting model is trained using training data. First, the neural network type of the hidden layer in the ForecastNet series forecasting unit is selected. Second, the hyperparameter optimization method is selected. Finally, the forecast time scale and forecast time interval are determined, and the input and output data of the corresponding time period are fed into the ForecastNet forecasting model for training. Step 5. Use the DTR timing prediction model to predict DTR Through model training, a DTR prediction model for a particular route is obtained. Environmental forecast data for the area where the route is located is obtained from the NWP: ambient temperature, wind speed, wind direction, and sunshine intensity. The time scale and time interval of the data selection must be the same as those for the training model. The characteristic sequence of the meteorological environment input for a period of time in the future is input into the model to obtain the DTR prediction value for each period of the future. Step 6. Implement rolling forecast In order to realize the rolling prediction of DTR, the NWP prediction data of the period that has occurred and the actual DTR data are merged into the historical database and used as training data to participate in model training to complete the DTR prediction model for the next period.

Images