A high-voltage circuit breaker running-in method and system based on a DDPG network model

Abstract

The present application relates to the technical field of high-voltage circuit breaker running-in, in particular to a high-voltage circuit breaker running-in method and system based on a DDPG network model; comprising the steps of: obtaining historical running-in data, the historical running-in data including an operation parameter feature vector a_t of each t-time running-in determined by an operator according to experience, and a performance index parameter feature vector s_t after the running-in; establishing a data set with corresponding running-in data of multiple complete running-in processes; etc.; predicting the next optimal running-in operation parameter from the corresponding performance index of the previous running-in operation in the historical data by training the DDPG network model, so as to achieve the best running-in effect in the shortest running-in times and determine the number of times of the best running-in endpoint.

Description

Technical Field

[0001] This invention relates to the field of high-voltage circuit breaker break-in technology, specifically to a high-voltage circuit breaker break-in method and system based on the DDPG network model. Background Technology

[0002] High-voltage circuit breakers are the most important control and protection equipment in power systems, and their reliability directly affects the safe and stable operation of the power grid. Newly installed high-voltage circuit breakers require a certain number of break-in operations for their moving mechanisms and contact systems to eliminate machining burrs, adjust fit clearances, and ensure the contact surfaces reach a good condition, thereby ensuring their specified mechanical and electrical characteristics. Traditional break-in operations typically use fixed parameters (such as fixed operating voltage and operating time intervals) for two or even several hundred no-load opening and closing operations. This method has significant drawbacks: 1. Low efficiency: The break-in process is conservative, with a large deviation between the number of operations and actual needs, wasting time and resources and increasing mechanical wear on the equipment. 2. Poor results: Fixed parameters cannot adapt to the subtle differences between individual circuit breakers, making it difficult to achieve optimal break-in results. 3. Reliance on experience: Judging the end point of the break-in process mainly relies on the operator's experience, judging by measuring changes in mechanical characteristic parameters such as opening and closing time and speed, or listening to sounds. This lacks objective and quantitative standards, easily leading to under-break-in or over-break-in. In recent years, some studies have attempted to assess circuit breaker status by monitoring more state variables such as opening and closing times, vibration signals, and overtravel. However, how to dynamically and automatically adjust the operating parameters for the next break-in period based on this massive amount of high-dimensional state information, and ultimately achieve the optimal break-in state in the shortest possible time, remains a pressing technical challenge. Current technologies lack an intelligent decision-making system capable of self-learning and self-optimization to accomplish this complex task. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for optimizing the break-in operation parameters of high-voltage circuit breakers based on the DDPG network model. This solves the problem that the existing break-in technology relies too much on human experience and is difficult to automatically set different break-in operation parameters for each break-in operation based on the individual differences of different high-voltage circuit breakers, and to determine the break-in endpoint.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0005] A break-in method for high-voltage circuit breakers based on a DDPG network model includes the following steps:

[0006] Historical break-in data is acquired, including the operational parameter feature vector a_t for each break-in cycle (t) as judged by operators based on experience, and the performance index parameter feature vector s_t (TC_t, TO_t, E_t, OV_t) after that break-in cycle. TC_t is the closing time range of the t-th break-in cycle, TO_t is the opening time of the t-th break-in cycle, E_t is the high-frequency vibration energy after the t-th break-in cycle, and OV_t is the overtravel of the t-th break-in cycle. A dataset is established using the corresponding break-in data from multiple complete break-in cycles.

[0007] The historical break-in data dataset is used as training samples to train the DDPG network. During training, s_t-m, ..., s_t-2, s_t-1 are used as inputs and a_t is used as output. The reward weight r_t = w1 * (TC_{t-1} - TC_t) + w2 (TO_{t-1} - TO_t) + w3 * (E_{t-1} - E_t) + w4 * (OV_{t-1} - OV_t). When the maximum value of r_t in the most recent k operations is less than the preset termination threshold, the break-in end is determined and the current break-in process is terminated.

[0008] The trained DDPG network is used to predict the a_t of each operation of the high-voltage vacuum circuit breaker that needs to be run-in; and the end point of this run-in is automatically determined.

[0009] Furthermore, in the above-mentioned high-voltage circuit breaker break-in method based on the DDPG network model, both the Actor and Critic networks in the DDPG network adopt two fully connected layers, with 256 neurons in each layer.

[0010] Furthermore, in the above-mentioned high-voltage circuit breaker break-in method based on the DDPG network model, the historical step size in the DDPG network model is m=3, and the termination judgment window is k=3.

[0011] Furthermore, in the above-mentioned high-voltage circuit breaker break-in method based on the DDPG network model, the termination threshold is set to 0.015.

[0012] Furthermore, in the above-mentioned high-voltage circuit breaker break-in method based on the DDPG network model, the operating parameter feature vector a_t is (V1, V2, θ), where V1 is the closing voltage, V2 is the opening voltage, and θ is the operating time interval.

[0013] Furthermore, in the above-mentioned high-voltage circuit breaker break-in method based on the DDPG network model, the performance index parameter feature vector s_t (TC_t, TO_t, E_t, OV_t) after each break-in is used to determine whether the corresponding TC_t, TO_t, E_t, and OV_t exceed the preset threshold. If they do, an alarm is triggered and subsequent break-in operations are stopped.

[0014] Furthermore, in the above-mentioned high-voltage circuit breaker break-in method based on the DDPG network model, the performance index parameter feature vector s_t is normalized before training the DDPG network model; the operation parameter feature vector a_t is normalized before training the DDPG network model.

[0015] Furthermore, in the above-mentioned high-voltage circuit breaker break-in method based on the DDPG network model, during the first m break-in operations in the actual break-in operation, the data with the smallest r_3 value corresponding to the first 3 break-in operations in the historical data is directly adopted.

[0016] This invention also relates to a high-voltage circuit breaker break-in system based on a DDPG network model. Based on the aforementioned high-voltage circuit breaker break-in method based on the DDPG network model, the break-in system includes:

[0017] The historical break-in data acquisition module is used to acquire historical break-in data, which includes the operational parameter feature vector a_t for each break-in cycle t, judged by the operator based on experience, and the performance index parameter feature vector s_t (TC_t, TO_t, E_t, OV_t) after that break-in cycle. TC_t is the closing time range of the t-th break-in process, TO_t is the opening time of the t-th break-in process, E_t is the high-frequency vibration energy after the t-th break-in cycle, and OV_t is the overtravel of the t-th break-in process. A dataset is established using the corresponding break-in data from multiple complete break-in cycles.

[0018] DDPG Network Training Module: Used to train the DDPG network using historical break-in data as training samples. During training, s_t-m, ..., s_t-2, s_t-1 are inputs, and a_t is output. The reward weight r_t = w1 *(TC_{t-1} - TC_t) + w2 (TO_{t-1} - TO_t) + w3 * (E_{t-1} - E_t) + w4 * (OV_{t-1} - OV_t). When the maximum value of r_t in the most recent k operations is less than the preset termination threshold, the break-in period is determined to be over, and the current break-in process is terminated.

[0019] Operating parameter prediction module: Used to predict the a_t of each operation of the high-voltage vacuum circuit breaker that needs to be run-in using a trained DDPG network; and automatically determine the end point of this run-in.

[0020] The beneficial effects of this invention are as follows: In the break-in operation of a high-voltage circuit breaker, the closing voltage, opening voltage, and operation time interval in each break-in operation are adjusted according to the change patterns of the corresponding performance indicators (closing time, opening time, high-frequency vibration energy, and overtravel) of the high-voltage circuit breaker after the previous few break-in operations. By training the DDPG network model, the optimal break-in operation parameters for the next operation are predicted based on the corresponding performance indicators of the previous few break-in operations from historical data. Thus, the optimal break-in effect can be achieved with the fewest number of break-in operations, and the optimal number of break-in endpoints can be determined. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a high-voltage circuit breaker break-in method based on a DDPG network model, which is a specific embodiment of the present invention. Detailed Implementation

[0022] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0023] Reference Figure 1 The specific embodiments of the present invention relate to a high-voltage circuit breaker break-in method based on a DDPG network model;

[0024] Includes the following steps:

[0025] Historical break-in data is acquired, including the operational parameter feature vector a_t for each break-in cycle (t) as judged by operators based on experience, and the performance index parameter feature vector s_t (TC_t, TO_t, E_t, OV_t) after that break-in cycle. TC_t is the closing time range of the t-th break-in cycle, TO_t is the opening time of the t-th break-in cycle, E_t is the high-frequency vibration energy after the t-th break-in cycle, and OV_t is the overtravel of the t-th break-in cycle. A dataset is established using the corresponding break-in data from multiple complete break-in cycles.

[0026] The historical break-in data dataset is used as training samples to train the DDPG network. During training, s_t-m, ..., s_t-2, s_t-1 are used as inputs and a_t is used as output. The reward weight r_t = w1 * (TC_{t-1} - TC_t) + w2 (TO_{t-1} - TO_t) + w3 * (E_{t-1} - E_t) + w4 * (OV_{t-1} - OV_t). When the maximum value of r_t in the most recent k operations is less than the preset termination threshold, the break-in end is determined and the current break-in process is terminated.

[0027] The trained DDPG network is used to predict the a_t of each operation of the high-voltage vacuum circuit breaker that needs to be run-in; and the end point of this run-in is automatically determined.

[0028] Specifically:

[0029] The test object was the ZW32-12 type indoor high-voltage vacuum circuit breaker.

[0030] Set the range of operating parameters, including:

[0031] Operating voltage: The standard operating voltage is DC24V, specifically including closing voltage and opening voltage. The closing voltage V1 ranges from 20.4 to 26.4V; the opening voltage V2 ranges from 15.6 to 28.8V.

[0032] The operation time interval θ ranges from 8 to 15 seconds;

[0033] The corresponding operating parameters for each t-time break-in operation are selected from the above range and normalized to form the feature vector a_t(V1, V2, θ).

[0034] Performance metrics range, including:

[0035] The closing time TC_t ranges from 25 to 60 ms;

[0036] The opening time TO_t ranges from 18 to 45 ms.

[0037] The high-frequency vibration energy E_t ranges from 0 to 1500 a.u. This performance indicator is measured using an IEPE (Integrated Circuit Piezoelectric) accelerometer. The sensor is securely mounted in a critical location on the circuit breaker body, such as above the porcelain bushing in the arc-extinguishing chamber, using a steel magnetic base or beeswax bonding. The magnetic base must be flat and clean to ensure high-frequency response. A high-speed acquisition card (sampling rate at least 2.5 times the highest frequency of interest, i.e., ≥ 125 kS / s) is used to record the vibration time-domain signal. The original signal is bandpass filtered (e.g., retaining the 5 kHz - 50 kHz component) to eliminate interference from low-frequency mechanical vibration and noise. The filtered signal is then calculated. The most common method is to calculate its root mean square (RMS) value or sum of squares as a measure of vibration energy.

[0038] The overtravel range (OV_t) is 2.5-5mm. This performance indicator uses a laser displacement sensor, such as the Keyence IL series. The sensor is fixed on a stable bracket, aligned with the reflective surface of the moving contact, ensuring the laser path is not obstructed during measurement. Overtravel refers to the distance the moving contact of the circuit breaker continues to move after contacting the stationary contact. It ensures that the contacts maintain sufficient contact pressure under electrodynamic and thermal effects.

[0039] After each t-time break-in operation, the corresponding performance parameters are measured by the above sensors, and after normalization, a performance parameter vector s_t (TC_t, TO_t, E_t, OV_t) is obtained. Each time, it is determined whether the corresponding performance in the corresponding performance vector is within the preset threshold. If it is, the next break-in operation is performed; otherwise, an alarm is triggered and the subsequent break-in operation is stopped.

[0040] Set the hyperparameters as follows: historical step size m=3, termination judgment window k=3, termination threshold 0.015, reward weights w1=1.0, w2=0.8, w3=1.2;

[0041] Reward weight r_t = w1 * (TC_{t-1} - TC_t) + w2 (TO_{t-1} - TO_t) + w3 *(E_{t-1} - E_t) + w4 * (OV_{t-1} - OV_t);

[0042] If the maximum value of r_t is less than the preset threshold in the most recent k operations (k=3), then the break-in period is determined to be over and the current break-in process is terminated.

[0043] Initialize the DDPG agent. Both the Actor and Critic networks use two fully connected layers, with 256 neurons in each layer.

[0044] A random strategy was used to run-in multiple batches of ZW32-12 indoor high-voltage vacuum circuit breakers. After each t-cycle break-in operation, the corresponding s_t was obtained through sensors. Then, experienced break-in operators, based on historical data, determined the feature vector a_t+1 for the next break-in operation. The corresponding s_t+1 and r_t+1 were then obtained through sensors. The operators then determined the feature vector a_t+2 for the next break-in operation, and so on, until a_n and s_n. The break-in operation ended when the corresponding r_n was less than a preset threshold. The recorded process data was used as historical data and as training samples. The training and test sets were divided in an 8:2 ratio to train the DDPG network. During training, s_t-3, s_t-2, and s_t-1 were used as inputs, and a_t was used as the output.

[0045] A trained DDPG network is used to predict the a_t parameter for each operation of the indoor high-voltage vacuum circuit breaker requiring break-in. For the first three break-in operations, the data with the smallest r_3 value from the previous three break-in operations in the historical data is directly used to obtain a_1, a_2, and a_3. From a_1, a_2, and a_3, s_1, s_2, and s_3 are obtained. Then, by inputting s_1, s_2, and s_3 into the trained DDPG network model, a_4 is output. The fourth break-in operation is performed using the a_4 parameter to obtain s_4, and r_4 is calculated accordingly. Then, s_2, s_3, and s_4 are input into the trained DDPG network model to output a_5, and so on, until after the nth break-in operation, when r_n is less than a preset threshold, the break-in operation is stopped.

[0046] After each actual break-in operation is completed, the break-in process is used as a new dataset to update the historical data and train and update the DDPG network.

[0047] This invention also relates to a high-voltage circuit breaker break-in system based on a DDPG network model. Based on the aforementioned high-voltage circuit breaker break-in method based on the DDPG network model, the break-in system includes:

[0048] The historical break-in data acquisition module is used to acquire historical break-in data, which includes the operational parameter feature vector a_t for each break-in cycle t, judged by the operator based on experience, and the performance index parameter feature vector s_t (TC_t, TO_t, E_t, OV_t) after that break-in cycle. TC_t is the closing time range of the t-th break-in process, TO_t is the opening time of the t-th break-in process, E_t is the high-frequency vibration energy after the t-th break-in cycle, and OV_t is the overtravel of the t-th break-in process. A dataset is established using the corresponding break-in data from multiple complete break-in cycles.

[0049] DDPG Network Training Module: Used to train the DDPG network using historical break-in data as training samples. During training, s_t-m, ..., s_t-2, s_t-1 are inputs, and a_t is output. The reward weight r_t = w1 *(TC_{t-1} - TC_t) + w2 (TO_{t-1} - TO_t) + w3 * (E_{t-1} - E_t) + w4 * (OV_{t-1} - OV_t). When the maximum value of r_t in the most recent k operations is less than the preset termination threshold, the break-in period is determined to be over, and the current break-in process is terminated.

[0050] Operating parameter prediction module: Used to predict the a_t of each operation of the high-voltage vacuum circuit breaker that needs to be run-in using a trained DDPG network; and automatically determine the end point of this run-in.

[0051] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A high-voltage circuit breaker running-in method based on a DDPG network model, characterized in that, The method comprises the steps of: acquiring historical grinding data, the historical grinding data including an operation parameter feature vector a_t of the tth grinding operation determined by an operator according to experience and a performance index parameter feature vector s_t (TC_t, TO_t, E_t, OV_t) after the grinding operation, TC_t being a closing time of the tth grinding operation, TO_t being an opening time of the tth grinding operation, E_t being a high-frequency vibration energy after the tth grinding operation, and OV_t being an overstroke of the tth grinding operation; and establishing a data set by using corresponding grinding data of multiple complete grinding operations; wherein the operation parameter feature vector a_t is (V1, V2, θ), V1 being a closing voltage, V2 being an opening voltage, and θ being an operation time interval; The dataset of historical running-in data is taken as a training sample to train the DDPG network, and in the training process, st-m,..., st-2, st-1 are taken as inputs, and at is taken as output; a reward weight r t = w1 (TC_{t-1} - TC_t) + w2 (TO_{t-1} - TO_t) + w3 (E_{t-1} - E_t) + w4 (OV_{t-1} - OV_t) is taken as output; when the maximum value of r t in the last k operations is less than a preset termination threshold, the running-in end point is determined, and the present running-in process is terminated; wherein m is a historical step length; using the trained DDPG network to predict an operation parameter feature vector of each operation of the high-voltage vacuum circuit breaker that needs to be ground, and automatically determining an end point of the present grinding.

2. The DDPG network model-based high-voltage circuit breaker running-in method according to claim 1, characterized in that, In the DDPG network, both the Actor and the Critic network use two fully connected layers, each layer having 256 neurons.

3. The DDPG network model-based high-voltage circuit breaker wear-in method according to claim 1, characterized in that, In the DDPG network model, m = 3 and k = 3.

4. The DDPG network model-based high-voltage circuit breaker wear-in method of claim 1, wherein, The termination threshold is set to 0.

015.

5. The DDPG network model-based high-voltage circuit breaker wear-in method according to claim 1, wherein, After each grinding operation, the performance index parameter feature vector s_t (TC_t, TO_t, E_t, OV_t) is used to determine whether TC_t, TO_t, E_t, and OV_t exceed the preset threshold, and if so, an alarm is given and the subsequent grinding operation is stopped.

6. The DDPG network model-based high-voltage circuit breaker wear-in method according to claim 1, characterized in that, The performance index parameter feature vector s_t is normalized before the DDPG network model is trained, and the operation parameter feature vector a_t is normalized before the DDPG network model is trained.

7. The DDPG network model-based high-voltage circuit breaker wear-in method according to claim 1, characterized in that, In the first m grinding operations of the actual grinding operation, the corresponding r_3 value of the first three corresponding grinding operations in the historical data is directly used.

8. A high-voltage circuit breaker running-in system based on a DDPG network model, characterized in that, The high-voltage circuit breaker grinding method based on the DDPG network model according to any one of claims 1 to 7, wherein the grinding system comprises: a historical grinding data acquisition module configured to acquire historical grinding data, the historical grinding data including an operation parameter feature vector a_t of the tth grinding operation determined by an operator according to experience and a performance index parameter feature vector s_t (TC_t, TO_t, E_t, OV_t) after the grinding operation, TC_t being a closing time of the tth grinding operation, TO_t being an opening time of the tth grinding operation, E_t being a high-frequency vibration energy after the tth grinding operation, and OV_t being an overstroke of the tth grinding operation; and a data set is established by using corresponding grinding data of multiple complete grinding operations; The DDPG network training module is configured to take a data set of historical running-in data as a training sample to train the DDPG network, and in the training process, take st-m,..., st-2, st-1 as input and take a_t as output; a reward weight r_t = w1 (TC_{t-1} - TC_t) + w2 (TO_{t-1} - TO_t) + w3 (E_{t-1} - E_t) + w4 (OV_{t-1} -OV_t); when the maximum value of r_t in the last k operations is less than a preset termination threshold, the running-in endpoint is determined, and the current running-in process is terminated. an operation parameter prediction module configured to use the trained DDPG network to predict an operation parameter feature vector of each operation of the high-voltage vacuum circuit breaker that needs to be ground, and automatically determine an end point of the present grinding.

Images