Power transmission line construction equipment parallel control method and device based on Bi-LSTM and DDPG algorithms

Abstract

The invention discloses a power transmission line construction equipment parallel control method and device based on Bi-LSTM and DDPG algorithms, and the method comprises the steps: optimizing the weight of a time sequence attention mechanism based on a differential evolution algorithm of a stage mutation strategy, and constructing a Bi-LSTM network model through a feature attention mechanism and the optimized time sequence attention mechanism; inputting the construction equipment state data into the Bi-LSTM network model to obtain a construction equipment state prediction result; and improving a DDPG algorithm based on priority post experience playback, constructing a decision control model by using the improved DDPG algorithm, and inputting a construction equipment state prediction result into the decision control model to obtain an optimal control strategy of the power transmission line construction equipment. According to the method, more accurate construction equipment state prediction is provided through the improved Bi-LSTM network model, and the decision control model constructed by using the improved DDPG algorithm makes a more optimized control strategy by using a state prediction result in combination with deep learning and a strategy gradient method, so that the solution of an optimal decision is realized.

Description

Technical Field

[0001] The present invention relates to the field of digital control of power transmission lines, and in particular to a method and device for parallel control of power transmission line construction equipment based on Bi-LSTM and DDPG algorithms. Background Art

[0002] With the rapid development of transmission lines, transmission line construction equipment faces severe challenges in terms of control efficiency, stability and flexibility. Specifically, the increasingly complex power transmission and distribution systems have put forward higher requirements for the real-time, precision and intelligence of control. However, traditional control methods based on static models are difficult to adapt to the frequent changes of such complex dynamic systems. In addition, the currently widely used timing analysis and optimization control algorithms also have certain limitations when processing complex data of large-scale power grids. In order to meet the above challenges, it is urgent to introduce cutting-edge technologies such as digital twins, deep learning and reinforcement learning to build a digital intelligent transmission network and realize efficient, stable and flexible control of line construction equipment.

[0003] In the existing technology, the document "Application Status and Development Prospect of Digital Twin Technology in State Assessment of Power Transmission and Transformation Equipment" proposes to deeply integrate digital twins with state assessment technology, and build digital twins of power transmission and transformation equipment based on technologies such as big data analysis and data mining, which can realize differentiated prediction and evaluation of equipment status and accurate and intelligent diagnosis of faults, and promote the operation and maintenance management of power transmission and transformation equipment into the intelligent era. The document "Discussion on Intelligent Inspection System of Transmission Lines Based on Digital Twin Technology" proposes a digital twin model of the transmission line inspection system, which realizes panoramic monitoring and state assessment of transmission line equipment, as well as online real-time monitoring and early warning of the visual state of line external failure scenes through a digital twin management and control platform; however, the existing digital twin system mainly plays the role of assisting judgment and decision-making, rather than directly generating control outputs, and needs to rely on manual intervention, resulting in limited actual automatic control capabilities. The document "Parallel Control of Thermal Power Units for Large-Scale New Energy Consumption" proposes a parallel control system for thermal power units. Based on digital twins and parallel intelligence theory, the system performs parallel execution and feedback correction of the manual control system and the actual system to significantly improve the flexible and intelligent adjustment capabilities of the units. The paper "Construction of Digital Twin of Excavation Face and Parallel Intelligent Control Method" proposes a parallel intelligent control architecture for excavation face based on digital twin and physical system. This architecture realizes the interactive integration of digital and physical excavation process by constructing digital twin of excavation face, so that advanced digital decision-making can assist and guide physical excavation control, and achieve the goal of digital twin and physical entity to move forward in coordination. The paper "Short-term Power Load Forecasting Combining Improved Residual Network and Bi-LSTM" proposes a bidirectional LSTM (Bi-LSTM) network that can simultaneously utilize forward and reverse context information, thereby enhancing the feature representation and association modeling capabilities of LSTM. However, when Bi-LSTM processes long sequences, intermediate information may still be attenuated or lost, and the model's interpretability is insufficient. In addition, the DDPG network parameters are obtained by random initialization, which limits the convergence ability of the initial agent parameters. Summary of the invention

[0004] The purpose of the present invention is to overcome the above-mentioned defects and problems existing in the prior art, and to provide a parallel control method and device for power transmission line construction equipment based on Bi-LSTM and DDPG algorithms. The improved Bi-LSTM network model provides a more accurate prediction of the state of construction equipment, and the decision control model constructed by the improved DDPG algorithm utilizes the state prediction results, combines deep learning and policy gradient methods, and formulates a more optimized control strategy, thereby achieving the solution of the optimal decision.

[0005] To achieve the above objectives, the technical solution of the present invention is:

[0006] A parallel control method for power transmission line construction equipment based on Bi-LSTM and DDPG algorithms, comprising:

[0007] The differential evolution algorithm based on the stage mutation strategy optimizes the weight of the temporal attention mechanism. A Bi-LSTM network model including input layer, feature attention layer, Bi-LSTM network layer, temporal attention layer, fully connected layer and output layer is constructed through the feature attention mechanism and the optimized temporal attention mechanism. The collected construction equipment status data is input into the Bi-LSTM network model to obtain the construction equipment status prediction result.

[0008] The DDPG algorithm is improved based on the post-experience playback of priority, and the decision control model is constructed using the improved DDPG algorithm. The prediction results of the construction equipment status are input into the decision control model to obtain the optimal control strategy.

[0009] The transmission line construction equipment is controlled through the optimal control strategy.

[0010] Based on digital twin technology, a digital twin model of transmission line construction equipment is constructed. The digital twin model includes a Bi-LSTM network model and a decision-making control model.

[0011] Based on parallel control theory, a parallel intelligent control model between the digital twin model and the transmission line equipment is constructed by adopting state data bidirectional synchronization technology and control instruction mapping technology.

[0012] The differential evolution algorithm based on the stage mutation strategy includes:

[0013] Initialize the basic parameters of the algorithm, including the population size N P , variable dimension L, number of iterations N I , initialization of mutation factor F and crossover ratio CR; the population is initialized using the following formula:

[0014]

[0015] In the formula, represents the optimization variable of the jth dimension of the i-th individual in the initial population, that is, the weight of the time-series attention mechanism; rand(0, 1) represents a random number that follows a uniform distribution in the interval (0, 1);

[0016] The evolutionary stage of the algorithm is evaluated based on the average Euclidean distance of the population:

[0017]

[0018] In the formula, is the average Euclidean distance of the g-th generation population; is the optimized variable of the jth dimension of the i-th individual in the g-th generation population; The best individual in the g-generation population;

[0019] calculate

[0020]

[0021] In the formula, is the normalized average Euclidean distance of the g-th generation population; d max is the maximum value of the average Euclidean distance of the population; d min is the minimum value of the average Euclidean distance of the population;

[0022] according to Estimate the current population stage:

[0023]

[0024] In the formula, ST represents the evolutionary stage of the population; S1, S2, and S3 represent the first, second, and third levels respectively; d c is the stage control factor;

[0025] Use the following formula to update the selection probability of the mutation strategy in the current stage:

[0026]

[0027] In the formula, p k is the selection probability of k mutation strategy; NS k The number of times the test individuals generated by the k mutation strategy in the current generation successfully enter the next generation;

[0028] The mutant individual and the target individual are mixed with parameters to generate the test individual, specifically:

[0029]

[0030] In the formula, represents the optimized variable of the jth dimension of the i-th mutant individual in the g-generation population; represents the optimization variable of the jth dimension of the i-th target individual in the g-generation population; represents the optimized variable of the jth dimension of the i-th test individual in the g+1 generation population after the crossover operation; j rand is a random integer, j rand ∈[1, 2, ..., L];

[0031] Use the greedy strategy to select the next generation, as shown below:

[0032]

[0033] In the formula, is the i-th individual of generation g+1; is the i-th individual of generation g; Represents the i individuals in the g+1 generation population after the crossover operation.

[0034] The input feature extraction link of the Bi-LSTM network layer is improved based on the feature attention mechanism:

[0035]

[0036] In the formula, q t =[q 1，t ,q 2，t , ..., q K，t ] is the feature attention weight coefficient vector at time t; W q is the trainable weight matrix; x t =[x 1，t , x 2，t , ..., x K，t ] is the input feature vector; v q is the bias vector of the feature attention weight; σ(·) is the Sigmoid activation function; A t =[A 1，t , A 2，t ..., A K，t ] is the feature attention weight vector, K is the number of features; x′ t is the weighted input vector;

[0037] The output layer of the Bi-LSTM network layer is improved based on the optimized temporal attention mechanism:

[0038]

[0039] In the formula, p t =[p 1，t , p 2，t , ..., p N，t ] is the temporal attention weight coefficient vector at time t; W r is the trainable weight matrix; H t =[H 1，t , H 2，t , ..., H N，t ] is the hidden layer state of the LSTM network unit; b r is the bias vector of the temporal attention weight; ReLU(·) is the activation function; B t =[B 1，t , B 2，t , ..., B N，t ] is the temporal attention weight vector, N is the length of the input sequence time window; H′t is the weighted input vector; is the matrix product.

[0040] The priority-based post-experience review includes:

[0041] Calculate the mean D of the trajectory target distance traj for:

[0042] D traj (τ) = D traj (ag0, ag1, ..., ag T )=arg max(d t );

[0043] Where τ is the trajectory; d t is the spatial distance between the desired target g and the achieved target ag at time t;

[0044] Calculate the priority value p of the trajectory i for:

[0045] p i =|-k*D trai +b|;

[0046] In the formula, k is the slope of the inverse proportional function; b is the intercept of the inverse proportional function;

[0047] Calculate the priority sampling probability p(τ) of the trajectory being replayed i )for:

[0048]

[0049] Where N is the total number of trajectories in the experience pool;

[0050] Recalculate the reward as:

[0051]

[0052] Where r1 is the reward calculated based on the spatial distance between the expected target g and the achieved target ag; β is the reward control factor; r2 is the reward obtained at time t+1; r ε is a constant; d t+1 is the spatial distance between the desired target g and the achieved target ag at time t+1;

[0053] The reward for hindsight experience is:

[0054]

[0055] Where r′ is the binary reward of the subsequent experience; r κ is the critical reward value.

[0056] The specific process of the improved DDPG algorithm is as follows:

[0057] (1) Initialize DDQN training network parameters θ and update target network parameters Then initialize the experience revisit pool D, iteration round e and maximum number of iterations E;

[0058] (2) Sampling target state g and initialization state s0;

[0059] (3) Environmental interaction:

[0060] a. Collect environmental status s t , select action a based on the current training network and ε-greedy strategy t ;

[0061]

[0062] Where ε is the exploration rate; λ is a randomly generated number between 0 and 1; A is the action set;

[0063] b. Execute action a t , and collect the environmental state s obtained after the digital twin performs the action t+1 ;

[0064] (4) Calculation of priority sampling;

[0065] a. Calculate the spatial distance between the desired target g and the achieved target ag;

[0066] b. Calculate the mean of the trajectory target distance;

[0067] c. Calculate the priority value of the trajectory;

[0068] d. Calculate the priority sampling probability of the trajectory being replayed;

[0069] (5) Post-experience review:

[0070] a. Storage transfer samples (s t ‖g,a t , r t ,s t+1 ‖g,p,D traj to the experience pool;

[0071] b. Sample trajectories from the experience pool according to the priority sampling probability of the trajectory being replayed;

[0072] c. Then sample from the trajectory (s t , a t ,s t+1 ), sample the virtual target state g′∈{s t+1, ..., s T-1};

[0073] d. Recalculate the reward for the transferred sample;

[0074] e. Storage of post-transfer samples (s t ||g′,a t , r′ t ,s t+1 ||g′,p,D traj ) to the experience pool D;

[0075] (6) Sample from the experience pool D to train the network and calculate the Q value of the training network and the target network Value, and update the training network parameters θ;

[0076]

[0077]

[0078] In the formula, is the target value of the sample extracted from the experience replay pool, μ and θ are the learning rate and training network parameters respectively; Q(·) and are the state-action functions of the training network and the target network respectively;

[0079] (7) Copy the training network parameters θ at a fixed frequency to update the target network parameters

[0080] (8) If t ≥ T, continue with the following steps, otherwise t = t + 1, return to step (3);

[0081] (9) If e≥E, exit the loop; otherwise, e=e+1 and return to step (3).

[0082] The improved DDPG algorithm is used to construct a decision control model, including:

[0083] State space: the state of the transmission line construction equipment, including the predicted values ​​of current, voltage and load; the predicted values ​​of the transmission line construction equipment parameters and the equipment operability constraints;

[0084] Action Space:

[0085] A t ={δ t , ΔP f};

[0086] In the formula, δ t is the switching action of the circuit breaker and disconnector; ΔP t is the change in power of construction equipment;

[0087] Reward function:

[0088]

[0089] In the formula, w i and w j are the weights of the indicator reward item and penalty item respectively; r i and p j are the i-th reward item value and the j-th penalty item value respectively.

[0090] A parallel control device for power transmission line construction equipment based on Bi-LSTM and DDPG algorithms, the device is applied to the above-mentioned method, the device comprising:

[0091] The construction equipment status prediction result acquisition module is used to optimize the weight of the temporal attention mechanism based on the differential evolution algorithm of the stage mutation strategy, and construct a Bi-LSTM network model including an input layer, a feature attention layer, a Bi-LSTM network layer, a temporal attention layer, a fully connected layer and an output layer through the feature attention mechanism and the optimized temporal attention mechanism; the collected construction equipment status data is input into the Bi-LSTM network model to obtain the construction equipment status prediction result;

[0092] The optimal control strategy acquisition module is used to improve the DDPG algorithm based on the post-experience playback of priority, build a decision control model using the improved DDPG algorithm, and input the construction equipment state prediction results into the decision control model to obtain the optimal control strategy;

[0093] The construction equipment control module is used to control the transmission line construction equipment through the optimal control strategy.

[0094] A parallel control device for power transmission line construction equipment based on Bi-LSTM and DDPG algorithms, comprising a memory and a processor;

[0095] The memory is used to store computer program code and transmit the computer program code to the processor;

[0096] The processor is used to execute the method according to the instructions in the computer program code.

[0097] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described above is implemented.

[0098] Compared with the prior art, the present invention has the following beneficial effects:

[0099] In a parallel control method and device for power transmission line construction equipment based on Bi-LSTM and DDPG algorithms of the present invention, the method first inputs the collected construction equipment state data into an improved Bi-LSTM network model to obtain a construction equipment state prediction result, then inputs the construction equipment state prediction result into a decision control model constructed using an improved DDPG algorithm to obtain an optimal control strategy, and then controls the power transmission line construction equipment through the optimal control strategy. The present invention provides a more accurate prediction of the state of construction equipment through an improved Bi-LSTM network model, and the decision control model constructed using the improved DDPG algorithm uses the state prediction result, combined with deep learning and policy gradient methods, to formulate a more optimized control strategy, thereby achieving the solution of the optimal decision, so as to accurately and efficiently solve the control problem of power transmission line construction equipment. The present invention can improve the control performance and anti-interference stability of power transmission line construction equipment, and promote the construction of digital intelligent power transmission networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] Figure 1 It is a flow chart of the parallel control method of power transmission line construction equipment based on Bi-LSTM and DDPG algorithms of the present invention.

[0101] Figure 2 It is a digital twin virtual model of the transmission line construction equipment proposed in the present invention.

[0102] Figure 3 This is the architecture diagram of the improved Bi-LSTM network model proposed in the present invention.

[0103] Figure 4 It is a schematic diagram of voltage control results of a case set in an embodiment of the present invention.

[0104] Figure 5 It is a schematic diagram of frequency control results of a case set in an embodiment of the present invention.

[0105] Figure 6 It is a schematic diagram of indicator parameter values ​​for setting cases in an embodiment of the present invention.

[0106] Figure 7 It is a comparison chart of the convergence of the algorithm proposed in the present invention and other algorithms.

[0107] Figure 8 It is a structural block diagram of a parallel control device for power transmission line construction equipment based on Bi-LSTM and DDPG algorithms of the present invention.

[0108] Fig. 9 It is a structural block diagram of a parallel control device for power transmission line construction equipment based on Bi-LSTM and DDPG algorithms of the present invention. DETAILED DESCRIPTION

[0109] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0110] See also Figure 1 , a parallel control method for power transmission line construction equipment based on Bi-LSTM and DDPG algorithm, comprising:

[0111] S1. The differential evolution algorithm (DE) based on the stage mutation strategy optimizes the weight of the temporal attention mechanism. A Bi-LSTM network model including input layer, feature attention layer, Bi-LSTM network layer, temporal attention layer, fully connected layer and output layer is constructed through the feature attention mechanism and the optimized temporal attention mechanism. The collected construction equipment status data is input into the Bi-LSTM network model to obtain the construction equipment status prediction result.

[0112] S2. The DDPG algorithm is improved based on priority hindsight experience replay (HPER), a decision control model is constructed using the improved DDPG algorithm, and the construction equipment status prediction results are input into the decision control model to obtain the optimal control strategy.

[0113] S3. Control the transmission line construction equipment through the optimal control strategy.

[0114] With the widespread application of new technologies such as wireless sensor networks, 5G, drone patrol and visualization in the field of power transmission lines, the efficiency of power transmission line construction and operation and maintenance has been significantly improved, laying the foundation for the intelligentization of power transmission lines. Facing the demand for comprehensive perception and intelligent analysis of the construction status of transmission lines, the rapid development of digital twin technology has provided new ideas for solving these problems. Based on the general digital twin architecture, this paper proposes a specific digital twin architecture suitable for power transmission line construction equipment. Figure 2 The architecture is based on the cloud, pipe, edge, and end structure of the power Internet of Things, and takes into account the unique characteristics of transmission line construction equipment. It mainly includes three key parts: perception layer, edge computing layer, and digital twin layer. The details are as follows:

[0115] a) Perception layer: mainly includes two types of equipment: construction equipment and collection equipment. Construction equipment includes rotary drilling rigs, hydraulic tension machines, hydraulic traction machines, winches, etc.; collection equipment includes measurement and control equipment, video monitoring equipment, and environmental monitoring equipment, which are used to collect and control data from transmission line construction equipment.

[0116] b) Edge computing layer: Edge computing devices include power smart terminals and aggregation gateways with edge computing capabilities. Smart terminals are responsible for access management, data analysis and collection, edge computing, and data forwarding to the cloud for construction equipment at the transmission line end. Its joint IoT platform enables connection and access to end devices, and provides IoT open services such as real-time data interaction, inspection strategy issuance, and real-time control. Edge computing realizes the front-end and back-end collaboration of data to ensure all-weather intelligent monitoring of the line. For environments with complex environments and limited communication capabilities, data is transmitted to the aggregation gateway through multi-hop aggregation, and then the gateway communicates with the cloud platform. This structure not only realizes efficient transmission and storage of data, but also ensures real-time monitoring of lines in complex environments.

[0117] c) Digital twin layer: The digital twin layer uses virtualization technology to build a virtual model that is highly corresponding to the actual physical system, and establishes a two-way interactive interface between digital and physical systems to reflect and simulate the state, behavior, and performance parameters of the physical system in real time.

[0118] The specific functions of the three key parts on the power transmission line construction equipment are as follows: A large amount of data originates from physical equipment, and data is collected through various acquisition devices in the perception layer. The data is first aggregated and analyzed in the edge computing layer, and then the edge computing device transmits the data to the cloud platform through optical fiber and wireless network. The cloud platform receives the data and passes it to the digital twin layer, and then integrates the data and simulates it through technologies such as modeling management, simulation services, and digital twin co-intelligence. Finally, human-computer interaction is realized through virtualization and visualization. Users can issue control instructions through the visual interactive interface and finally send them to the physical equipment of the power transmission line construction in the perception layer to realize intelligent control of the equipment.

[0119] Based on parallel control theory and using state data bidirectional synchronization and control instruction mapping technology, a parallel intelligent control architecture that integrates data and objects is constructed.

[0120] In the smart grid environment, in order to realize the coordinated parallel control of the digital-physical integration of transmission line construction equipment, the state data bidirectional synchronization technology and control instruction mapping technology play a key role. First, sensors and monitoring equipment are deployed to collect real-time data related to the state of construction equipment, such as equipment operating parameters such as current, voltage, power, and temperature. Then, the state data is transmitted to the digital twin model, and the internal state of the model is updated to accurately reflect the real-time state and environmental changes of the actual system. Secondly, the digital twin generates control instructions, and the control instructions generated by the digital twin model are accurately and effectively mapped to the actual physical system through the control instruction mapping technology. Then, the response of the actual physical system is monitored in real time, that is, the state changes of the actual transmission line construction equipment after implementation. These response data are used to verify the accuracy of the digital twin model and the effectiveness of the optimization control strategy. Finally, an online state detection and evaluation module is set to check the rationality of resource scheduling and load distribution of the physical system after receiving the control instructions of the digital twin. If the control efficiency is lower than expected, a new round of digital simulation prediction and control update is started to complete the dynamic adjustment of the physical system. The response data is then fed back to the digital twin model to adjust the model parameters, improve its fit to the actual system, and ensure that the model is continuously optimized to improve adaptability.

[0121] Furthermore, a digital twin model of transmission line construction equipment is constructed based on digital twin technology. The digital twin model includes a Bi-LSTM network model and a decision control model. Based on parallel control theory, a parallel intelligent control model between the digital twin model and the transmission line equipment is constructed by using state data bidirectional synchronization technology and control instruction mapping technology.

[0122] In another embodiment of the present application, a differential evolution algorithm based on a stage mutation strategy includes:

[0123] (1) Initialization

[0124] Initialize the basic parameters of the algorithm, including the population size N P , variable dimension L, number of iterations N I , initialization of mutation factor F and crossover ratio CR; the population is initialized using the following formula:

[0125]

[0126] In the formula, represents the optimization variable of the jth dimension of the i-th individual in the initial population, that is, the weight of the time-series attention mechanism; rand(0,1) represents a random number that follows a uniform distribution in the interval (0,1);

[0127] (2) Mutation

[0128] Considering the simplicity and complexity of the mean Euclidean distance (MED) calculation, the evolutionary stage of the algorithm is evaluated according to the mean Euclidean distance of the population:

[0129]

[0130] In the formula, is the average Euclidean distance of the g-th generation population; is the optimized variable of the jth dimension of the i-th individual in the g-th generation population; The best individual in the g-generation population;

[0131] calculate

[0132]

[0133] In the formula, is the normalized average Euclidean distance of the g-th generation population; d max is the maximum value of the average Euclidean distance of the population; d min is the minimum value of the average Euclidean distance of the population;

[0134] First, take the maximum value d of the population MED from the MED of the initial population max , and then the MED of each evolutionary population and d max Compare, if MED is greater than d max , then replace d max .

[0135] Finally, according to Estimate the current population stage:

[0136]

[0137] In the formula, ST represents the evolutionary stage of the population; S1, S2, and S3 represent the first, second, and third levels respectively; d c is the stage control factor;

[0138] In each stage, the selection probability of the first generation mutation strategy is initialized to 0.5, and then the selection probability of the mutation strategy in the current stage is updated using the following formula:

[0139]

[0140] In the formula, p k is the selection probability of k mutation strategy; NS k The number of times the test individuals generated by the k mutation strategy in the current generation successfully enter the next generation.

[0141] In order to avoid each mutation strategy being lost due to poor performance during the evolution process, a constant ξ = 0.01 is added to the selection probability of each mutation strategy.

[0142] In the evolutionary process, the stage of the current generation is first determined, and then the mutation strategy is selected using the roulette method according to the selection probability of each strategy in the corresponding stage to achieve adaptive selection of the mutation strategy.

[0143] (3) Cross

[0144] The mutant individual and the target individual are mixed with parameters to generate the test individual, specifically:

[0145]

[0146] In the formula, represents the optimized variable of the jth dimension of the i-th mutant individual in the g-generation population; represents the optimization variable of the jth dimension of the i-th target individual in the g-generation population; represents the optimized variable of the jth dimension of the i-th test individual in the g+1 generation population after the crossover operation; j rand is a random integer, j rand ∈[1, 2, ..., L].

[0147] The mutant individual refers to the new vector obtained after the individual vector is mutated. The target individual is the vector that has not been mutated. Most of them contain relevant heuristic experience that makes the target optimal. They are the high-quality individuals of the parent generation left after crossover selection. When the mixing condition is not met, is directly equal to By mixing mutant individuals with target individuals, it helps to introduce more different characteristics into the population, which guides the search direction to a certain extent, making it more likely that the newly generated test individuals will move in a better direction.

[0148] (4)Select

[0149] Use the greedy strategy to select the next generation, as shown below:

[0150]

[0151] In the formula, is the i-th individual of generation g+1; is the i-th individual of generation g; Represents the i individuals in the g+1 generation population after the crossover operation.

[0152] In another embodiment of the present application, see Figure 3The complete architecture of the improved Bi-LSTM network model proposed in this invention mainly includes input and output layers, feature attention layers, Bi-LSTM network layers, temporal attention layers and fully connected layers, and improves the LSTM network based on feature and temporal attention mechanisms. The input layer receives time series data; the feature attention layer automatically learns the association between input features through dynamic weight allocation; the Bi-LSTM layer learns the temporal information of the data; the temporal attention layer distinguishes the importance of different historical moments; and finally the output layer gives the prediction results.

[0153] First, a conventional Bi-LSTM model is constructed. Bi-LSTM consists of two stacked LSTM units, one LSTM unit processes the forward input sequence, and the other processes the reverse input sequence, and then connects the outputs of the two to capture both the forward order information and the reverse context information. Each LSTM unit contains: an input gate (determines the amount of input information passing through at the current moment), a forget gate (determines the degree of forgetting the previous state at the current moment), and an output gate (determines the amount of output information passing through at the current moment). The three gating mechanisms are defined as follows:

[0154]

[0155] Where i, f and o represent the input gate, forget gate and output gate respectively; ⊙ is the Hadamard product; W and b represent the weight matrix and bias vector of the network; x t 、h t and c t are the input and output vectors and memory units of the hidden layer in period t respectively.

[0156] Secondly, in order to solve the problem of feature association information loss in traditional correlation analysis methods, the present invention improves the input feature extraction link of the Bi-LSTM network layer based on the feature attention mechanism. This mechanism dynamically allocates attention weights and mines the correlation between input features and target features, thereby focusing attention on relevant features with large amounts of information and reducing or ignoring the impact of irrelevant features. The definition is as follows:

[0157]

[0158] In the formula, q t =[q 1，t ,q 2，t , ..., q K，t ] is the feature attention weight coefficient vector at time t; W q is the trainable weight matrix; x t =[x 1，t , x 2，t , ..., x K，t ] is the input feature vector; v qis the bias vector of the feature attention weight; σ(·) is the Sigmoid activation function; A t =[A 1，t , A 2，t , ..., A K，t ] is the feature attention weight vector, K is the number of features; x′ t is the weighted input vector;

[0159] Finally, in order to distinguish the different effects of each historical moment of the input sequence on the current prediction output, the present invention first improves the DE algorithm based on the stage mutation strategy. In this step, the weight of the temporal attention mechanism is optimized by the improved DE algorithm, and finally the output layer of the Bi-LSTM network layer is improved based on the optimized temporal attention mechanism. This mechanism dynamically assigns weights to different temporal information and autonomously learns the temporal association between data at each moment, thereby enhancing the information expression of key moments. The definition is as follows:

[0160]

[0161] In the formula, p t =[p 1，t , p 2，t , ..., p N，t ] is the temporal attention weight coefficient vector at time t; W r is the trainable weight matrix; H t =[H 1,t ,H 2,t ,...,H N,t ] is the hidden layer state of the LSTM network unit; b r is the bias vector of the temporal attention weight; ReLU(·) is the activation function; B t =[B 1，t , B 2，t , ..., B N，t ] is the temporal attention weight vector, N is the length of the input sequence time window; H′ t is the weighted input vector; is the matrix product.

[0162] In another embodiment of the present application, the DDPG algorithm achieves the ability to learn optimal control strategies in a continuous action space by combining deep learning with a policy gradient method. However, the traditional DDPG network has many problems in terms of training convergence speed and agent (control center) reward sparsity. In order to solve these problems, the present invention introduces the priority idea into the post-experience replay mechanism, and improves the DDPG algorithm based on priority post-experience replay (PHER). In addition, in response to the possible over-estimation problem in the re-estimation of post-experience rewards, the present invention reshapes the original reward.

[0163] First, define the trajectory-target distance function. The smaller the value of this function, the more helpful the action taken by the trajectory in achieving the target task is to get closer to the target task. The trajectory-target distance function mainly calculates the mean between the virtual target and the target task in one round, and calculates the priority probability of the trajectory based on this.

[0164] In one round, a complete trajectory τ can be represented by a tuple (S, A, P, R, γ). A trajectory consists of a series of consecutive states s t , where t represents the time step, t∈{0,1,...,T}. Each state s t ∈S also contains the state of the desired target, which represents a generalized subset of the desired state. The trajectory distance function D traj Depends on the target state τ = s1, a1, s2, a2, ..., s T-1 , a T-1 ,s T In the decision-making task, there are two goals, one is the desired goal g, and the other is the goal ag achieved after performing one step of operation. The agent expects to achieve the desired goal after performing several steps of operation. In the simulation environment of the task, the spatial positions of the two goals are represented by the Cartesian coordinate system, which are (x g ,y g , z g ) and (x ag ,y ag , z ag ), so the spatial distance between the two targets is:

[0165]

[0166] Calculate the mean D of the trajectory target distance traj for:

[0167] D traj (τ) = D traj (ag0, ag1, ..., ag T )=arg max(d t );

[0168] Where τ is the trajectory; d t is the spatial distance between the expected target g and the achieved target ag at time t;

[0169] Secondly, the spatial distance between the desired target g and the achieved target ag is calculated, and the trajectory experience with a small distance is given a higher priority probability for playback. At the beginning of each round, the agent adopts a random strategy to start exploring the environment. The sampled trajectories are stored in the experience pool. When the agent obtains a new trajectory, the distance is calculated using the spatial distance function of the target, and the distance value is stored in the experience pool together with the trajectory for subsequent priority sorting. Subsequently, the present invention uses a function with a negative slope to calculate the priority value of the trajectory. Because the function is a monotonically decreasing function, the trajectory distance value D traj (τ i ) is smaller, the higher its priority value is. Let the priority value of each trajectory p i for:

[0170] p i =|-k*D traj +b|;

[0171] In the formula, k is the slope of the inverse proportional function; b is the intercept of the inverse proportional function;

[0172] When sampling from the experience pool, the agent directly uses the proportion of the priority value of the trajectory in the sum of the priority values ​​of all trajectories as the priority sampling probability. This means that trajectories with high priority are more likely to be sampled and replayed. Mathematically, the priority sampling probability p(τ i )for:

[0173]

[0174] Where N is the total number of trajectories in the experience pool;

[0175] Finally, reshape the ex-post sample reward. In the DDPG algorithm with the goal as the condition, the reward of the sampled ex-post experience is recalculated based on the distance between the achieved goal and the expected goal in a time step, thereby alleviating the phenomenon of overestimation of the ex-post experience reward. The present invention changes the calculation method of the ex-post experience sample reward and recalculates the reward based on the distance, as follows:

[0176]

[0177] In the formula, r1 is the reward calculated based on the spatial distance between the expected target g and the achieved target ag. In the post-experience playback, the experience is a four-tuple (ag t ||g,a t , rt , ag t+1 ||g), where ag t Indicates that the goal has been achieved at time t, ag t+1 Indicates that the goal has been achieved at time t+1; β is the reward control factor; r2 is the reward obtained at time t+1, which can guide the agent closer and closer to the goal; r ε is a constant; d t is the expected goal g and the achieved goal ag at time t t The spatial distance between t+1 is the expected target g and the target ag achieved at time t+1 t+1 If d t+1 <d t , it means that the agent is approaching the goal, r2=r ε If d t+1 >d t , it means that the agent is moving away from the target, r2=-r ε .

[0178] The reward for hindsight experience is:

[0179]

[0180] Where r′ is the binary reward of the subsequent experience; r κ is the critical reward value.

[0181] The priority-based post-experience replay improved DDPG algorithm proposed in the present invention allows dynamic updating of experience data, and can more effectively utilize resources, improve sampling frequency, training efficiency and stability, which means that the algorithm can converge to a higher reward level faster in the same time.

[0182] Furthermore, the specific process of the improved DDPG algorithm is as follows:

[0183] (1) Initialize DDQN training network parameters θ and update target network parameters Then initialize the experience revisit pool D, iteration round e and maximum number of iterations E;

[0184] (2) Sampling the target state g and the initialization state S0;

[0185] (3) Environmental interaction:

[0186] a. Collect environmental status S t , select action a based on the current training network and ε-greedy strategy t ;

[0187]

[0188] Where ε is the exploration rate; λ is a randomly generated number between 0 and 1; A is the action set;

[0189] b. Execute action a t , and collect the environmental state S obtained after the digital twin performs the action t+1 ;

[0190] (4) Calculation of priority sampling;

[0191] a. Calculate the spatial distance between the desired target g and the achieved target ag;

[0192] b. Calculate the mean of the trajectory target distance;

[0193] c. Calculate the priority value of the trajectory;

[0194] d. Calculate the priority sampling probability of the trajectory being replayed;

[0195] (5) Post-experience review:

[0196] a. Storage transfer samples (s t ‖g,a t , r t ,s t+1 ‖g,p,D traj ) to the experience pool;

[0197] b. Sample trajectories from the experience pool according to the priority sampling probability of the trajectory being replayed;

[0198] c. Then sample from the trajectory (s t , a t ,s t+1 ), sample the virtual target state g′∈{s t+1 , ..., s T-1};

[0199] d. Recalculate the reward for the transferred sample;

[0200] e. Storage of post-transfer samples (s t ‖g′,a t , r′ t ,s t+1 ‖g′,p,D traj ) to the experience pool D;

[0201] (6) Sample from the experience pool D to train the network and calculate the Q value of the training network and the target network Value, and update the training network parameters θ;

[0202]

[0203] In the formula, is the target value of the sample extracted from the experience replay pool, μ and θ are the learning rate and training network parameters respectively; Q(·) and are the state-action functions of the training network and the target network respectively;

[0204] (7) Copy the training network parameters θ at a fixed frequency to update the target network parameters

[0205] (8) If t ≥ T, continue with the following steps, otherwise t = t + 1, return to step (3);

[0206] (9) If e≥E, exit the loop; otherwise, e=e+1 and return to step (3).

[0207] In another embodiment of the present application, in transmission line construction equipment, through reasonable problem modeling, state representation, action definition and reward design, the decision control model constructed by the improved DDPG algorithm can realize dynamic adjustment of the optimal control strategy in the actual control center.

[0208] State space: the state of the transmission line construction equipment, including the predicted values ​​of current, voltage and load; the predicted values ​​of the transmission line construction equipment parameters and the equipment operability constraints;

[0209] Action space: adjustment of device parameters, operation of switches, etc.; the definition of actions is within the acceptable range of the physical system and consistent with the actual operation;

[0210] A t ={δ t , ΔP t};

[0211] In the formula, δ t is the switching action of the circuit breaker and disconnector; ΔP t is the change in power of construction equipment;

[0212] Reward function:

[0213]

[0214] In the formula, w i and w j are the weights of the indicator reward item and penalty item respectively; r i and p j are the i-th reward item value and the j-th penalty item value respectively.

[0215] The reward function is consistent with the optimization goal of the control center, which is to reduce the scope of faults, reduce losses, and improve power supply reliability.

[0216] The present invention constructs a high-fidelity digital twin virtual model of transmission line construction equipment, and then based on parallel control theory, adopts state data bidirectional synchronization and control instruction mapping technology to construct a parallel intelligent control architecture of digital-physical integration, so that digital intelligent optimization decision-making assists physical control, and realizes intelligent collaboration and mutual intelligence between digital and physical systems. The differential evolution algorithm (DE) based on the stage mutation strategy optimizes the weight of the temporal attention mechanism, and on this basis, the Bi-LSTM network model is improved by feature attention and the optimized temporal attention mechanism, so as to dynamically capture the potential correlation between the temporal features of the construction equipment parameters and enhance the expression of key information; the Bi-LSTM network model is used to predict the state and performance of the construction equipment in the future. The DDPG algorithm is improved based on priority post-experience replay (HPER), and the decision control model is constructed using the improved DDPG algorithm. The results predicted by the Bi-LSTM network model are used as one of the inputs of the decision control model, such as the predicted values ​​of line construction equipment parameters, current and voltage, together with the current environmental state, for decision making by the intelligent agent (control center). The control center formulates the optimal control decision, such as adjusting the operating parameters or switch status of the construction equipment, based on the current state and the future state predicted by the Bi-LSTM network model, as well as the optimal strategy learned by the DDPG algorithm. In the actual construction process, the operation of the construction equipment is guided according to the decision of the control center to achieve efficient completion of the transmission line construction task. The improved Bi-LSTM network model provides a more accurate state prediction, and the improved DDPG decision control model uses these prediction results, combined with deep learning and policy gradient methods, to formulate more optimized control decisions, thereby achieving the solution of the optimal decision. The present invention can improve the control performance and anti-interference stability of transmission line construction equipment, and promote the construction of digital intelligent transmission networks.

[0217] In order to verify the effectiveness of the proposed parallel control of data-physical integration, a variety of comparison schemes are designed: 1. The proposed parallel control method; 2. Parallel control based only on improved DDPG decision; 3. Parallel control based only on improved Bi-LSTM prediction; 4. Traditional control method. And set a case: abnormal construction equipment. Set up 2 important construction equipment, whose load increases to 20%, and set adjacent splittable lines. Evaluate the node power supply pressure, and the control center formulates a priority power supply strategy for important equipment. According to the voltage and frequency stability analysis, a controlled splitting scheme is given to balance the load to improve power supply reliability. The case scenario simulates and verifies the priority equipment recovery strategy based on state assessment prediction, and responds to the voltage and frequency anomalies caused by abnormal equipment growth through priority power supply strategies and effective splitting schemes. When the power supply of the line on this side is insufficient due to the increase in equipment load, part of the equipment load is split in a controlled manner to the adjacent line for support, so as to ensure the stability of the voltage and frequency on this side and the power supply needs of important equipment. See. Figure 4 and Figure 5 It is the voltage and frequency control result of the case set up by the present invention. It can be seen from the figure that with the abnormal increase of equipment load, the voltage and frequency decrease synchronously. In order to ensure the power supply reliability, the method of the present invention restores the voltage and frequency to normal levels in the shortest possible time by reducing the energy consumption of equipment and the complementary energy of adjacent lines.

[0218] Furthermore, in order to intuitively compare the performance indicators of various solutions, the present invention is designed Figure 6 To represent the comparison results of the performance index parameters of the four schemes. It can be seen intuitively from the figure that in dynamic quality control tasks such as frequency and voltage regulation, the closed-loop parallel control driven by the digital twin achieves the lowest voltage / frequency stabilization time of 0.019 / 0.508 and the minimum voltage / frequency dynamic deviation maximum value of 0.021 / 0.18. Compared with other schemes, the maximum deviation of the voltage and frequency dynamic indicators is reduced by 40% and more than 17%, respectively, and the stable overshoot time is shortened by nearly 3.5% and more than 34%, which greatly reduces the dependence and regulation of the transmission line on stability during construction. In addition, the strategy proposed in the present invention enables the digital twin control instructions to the physical end. Compared with traditional static physical control, the scheme proposed in the present invention still achieves the best effect, the voltage qualification rate of important equipment is increased by 11%, the power supply satisfaction rate of important equipment is also increased by 9.8%, and the load balancing degree is also improved by 6%, which will significantly reduce the possibility of line overload.

[0219] Figure 7 This is a comparison of the convergence of the algorithm proposed in this invention and other algorithms. In order to verify the effectiveness of the improved algorithm proposed in this invention, the following comparison methods are set in the case scenario: 1. LSTM-DDPG; 2. LSTM-Improved DDPG; 3. Bi-LSTM-Improved DDPG; 4. The improved algorithm proposed in this invention. Figure 7 It can be seen that the performance of the improved algorithm proposed in the present invention is better than other methods. Specifically, the proposed improved algorithm reaches convergence after about 180 iterations, while comparison method 2, comparison method 3, and comparison method 4 reach convergence after about 380, 400, and 780 iterations, respectively, which shows that the speed of exploring the optimal strategy by the proposed method is increased by 52.6%, 55%, and 76.9%, respectively. Moreover, compared with other methods, the proposed method obtains the highest reward value of -1.03. These significantly reflect the effectiveness of the method proposed in the present invention.

[0220] See also Figure 8 The present invention also provides a parallel control device for power transmission line construction equipment based on Bi-LSTM and DDPG algorithms, which is applied to a parallel control method for power transmission line construction equipment based on Bi-LSTM and DDPG algorithms. The device comprises:

[0221] The construction equipment status prediction result acquisition module is used to optimize the weight of the temporal attention mechanism based on the differential evolution algorithm of the stage mutation strategy, and construct a Bi-LSTM network model including an input layer, a feature attention layer, a Bi-LSTM network layer, a temporal attention layer, a fully connected layer and an output layer through the feature attention mechanism and the optimized temporal attention mechanism; the collected construction equipment status data is input into the Bi-LSTM network model to obtain the construction equipment status prediction result;

[0222] The optimal control strategy acquisition module is used to improve the DDPG algorithm based on the post-experience playback of priority, build a decision control model using the improved DDPG algorithm, and input the construction equipment state prediction results into the decision control model to obtain the optimal control strategy;

[0223] The construction equipment control module is used to control the transmission line construction equipment through the optimal control strategy.

[0224] See also Fig. 9 ,The present invention also provides a parallel control device for power transmission line construction equipment based on Bi-LSTM and DDPG algorithms, comprising a memory and a processor;

[0225] The memory is used to store computer program code and transmit the computer program code to the processor;

[0226] The processor is used to execute a parallel control method for power transmission line construction equipment based on Bi-LSTM and DDPG algorithms according to instructions in the computer program code.

[0227] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, a parallel control method for power transmission line construction equipment based on Bi-LSTM and DDPG algorithms is implemented.

[0228] Generally speaking, the computer instructions for implementing the method of the present invention may be carried in any combination of one or more computer-readable storage media. Non-transitory computer-readable storage media may include any computer-readable media except for the signal itself that is temporarily propagating.

[0229] The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EKROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, device, or device.

[0230] Computer program code for performing the operation of the present invention may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages, in particular, Python suitable for neural network computing and platform frameworks based on TensorFlow, PyTorch, etc. may be used. The program code may be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer or to an external computer (e.g., using an Internet service provider to connect via the Internet) through any type of network, including a local area network (LAN) or a wide area network (WAN).

[0231] The above-mentioned devices and non-temporary computer-readable storage media can be found in the specific description of a parallel control method for transmission line construction equipment based on Bi-LSTM and DDPG algorithms and its beneficial effects, which will not be repeated here.

[0232] Although the embodiments of the present invention have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A parallel control method for power transmission line construction equipment based on Bi-LSTM and DDPG algorithm, characterized in that: include: The differential evolution algorithm based on the stage mutation strategy optimizes the weight of the temporal attention mechanism. A Bi-LSTM network model including input layer, feature attention layer, Bi-LSTM network layer, temporal attention layer, fully connected layer and output layer is constructed through the feature attention mechanism and the optimized temporal attention mechanism. The collected construction equipment status data is input into the Bi-LSTM network model to obtain the construction equipment status prediction result. The DDPG algorithm is improved based on the post-experience playback of priority, and the decision control model is constructed using the improved DDPG algorithm. The prediction results of the construction equipment status are input into the decision control model to obtain the optimal control strategy. The transmission line construction equipment is controlled through the optimal control strategy.

2. According to claim 1, a parallel control method for power transmission line construction equipment based on Bi-LSTM and DDPG algorithm is characterized in that: Based on digital twin technology, a digital twin model of transmission line construction equipment is constructed. The digital twin model includes a Bi-LSTM network model and a decision-making control model. Based on parallel control theory, a parallel intelligent control model between the digital twin model and the transmission line equipment is constructed by adopting state data bidirectional synchronization technology and control instruction mapping technology.

3. According to claim 1, a parallel control method for power transmission line construction equipment based on Bi-LSTM and DDPG algorithm is characterized in that: The differential evolution algorithm based on the stage mutation strategy includes: Initialize the basic parameters of the algorithm, including the population size N P , variable dimension L, number of iterations N I , initialization of mutation factor F and crossover ratio CR; the population is initialized using the following formula: In the formula, represents the optimization variable of the jth dimension of the i-th individual in the initial population, that is, the weight of the time-series attention mechanism; rand(0, 1) represents a random number that follows a uniform distribution in the interval (0, 1); The evolutionary stage of the algorithm is evaluated based on the average Euclidean distance of the population: In the formula, is the average Euclidean distance of the g-th generation population; is the optimized variable of the jth dimension of the i-th individual in the g-th generation population; The best individual in the g-generation population; calculate In the formula, is the normalized average Euclidean distance of the g-th generation population; d max is the maximum value of the average Euclidean distance of the population; d min is the minimum value of the average Euclidean distance of the population; according to Estimate the current population stage: In the formula, ST represents the evolutionary stage of the population; S1, S2, and S3 represent the first, second, and third levels respectively; d c is the stage control factor; Use the following formula to update the selection probability of the mutation strategy in the current stage: In the formula, p k is the selection probability of k mutation strategy; NS k The number of times the test individuals generated by the k mutation strategy in the current generation successfully enter the next generation; The mutant individual and the target individual are mixed with parameters to generate the test individual, specifically: In the formula, represents the optimized variable of the jth dimension of the i-th mutant individual in the g-generation population; represents the optimization variable of the jth dimension of the i-th target individual in the g-generation population; represents the optimized variable of the jth dimension of the i-th test individual in the g+1 generation population after the crossover operation; j rnad is a random integer, j rand ∈[1, 2, ..., L]; Use the greedy strategy to select the next generation, as shown below: In the formula, is the i-th individual of generation g+1; is the i-th individual of generation g; Represents the i individuals in the g+1 generation population after the crossover operation.

4. According to claim 1, a parallel control method for power transmission line construction equipment based on Bi-LSTM and DDPG algorithm is characterized in that: The input feature extraction link of the Bi-LSTM network layer is improved based on the feature attention mechanism: In the formula, q t =[q 1，t ,q 2，t ,...,q K，t ] is the feature attention weight coefficient vector at time t; W q is the trainable weight matrix; x t =[x 1，t , x 2，t , ..., x K，t ] is the input feature vector; v q is the bias vector of the feature attention weight; σ(·) is the sigmoid activation function; A t =|A 1，t , A 2，t ,...,A K，t | is the feature attention weight vector, K is the number of features; x′ t is the weighted input vector; The output layer of the Bi-LSTM network layer is improved based on the optimized temporal attention mechanism: In the formula, p t =[p 1：t , p 2，t , ..., p N，t ] is the temporal attention weight coefficient vector at time t; W r is the trainable weight matrix; H t =[H 1，t , H 2，t , ..., H N，t ] is the hidden layer state of the LSTM network unit; b r is the bias vector of the temporal attention weight; ReLU(·) is the activation function; B t =[B 1，t , B 2，t ,...,B N，t ] is the temporal attention weight vector, N is the length of the input sequence time window; H′ t is the weighted input vector; is the matrix product.

5. According to claim 1, a parallel control method for power transmission line construction equipment based on Bi-LSTM and DDPG algorithm is characterized in that: The priority-based post-experience review includes: Calculate the mean D of the trajectory target distance traj for: D traj (τ)=D traj (ag0,ag1,...,ag T )=arg max(d t ); Where τ is the trajectory; d t is the spatial distance between the expected target g and the achieved target ag at time t; Calculate the priority value p of the trajectory i for: p i =|-k*D traj +b|; In the formula, k is the slope of the inverse proportional function; b is the intercept of the inverse proportional function; Calculate the priority sampling probability p(τ) of the trajectory being replayed i )for: Where N is the total number of trajectories in the experience pool; Recalculate the reward as: Where r1 is the reward calculated based on the spatial distance between the expected target g and the achieved target ag; β is the reward control factor; r2 is the reward obtained at time t+1; r ε is a constant; d t+1 is the spatial distance between the expected target g and the achieved target ag at time t+1; The reward for hindsight experience is: Where r′ is the binary reward of the subsequent experience; r k is the critical reward value.

6. The method for parallel control of power transmission line construction equipment based on Bi-LSTM and DDPG algorithm according to claim 1, characterized in that: The specific process of the improved DDPG algorithm is as follows: (1) Initialize DDQN training network parameters θ and update target network parameters Then initialize the experience revisit pool D, iteration round e and maximum number of iterations E; (2) Sampling the target state g and the initialization state S0; (3) Environmental interaction: a. Collect environmental status S t , select action a based on the current training network and ε-greedy strategy t ; Where ε is the exploration rate; λ is a randomly generated number between 0 and 1; A is the action set; b. Execute action a t , and collect the environmental state S obtained after the digital twin performs the action t+1 ; (4) Calculation of priority sampling; a. Calculate the spatial distance between the desired target g and the achieved target ag; b. Calculate the mean of the trajectory target distance; c. Calculate the priority value of the trajectory; d. Calculate the priority sampling probability of the trajectory being replayed; (5) Post-experience review: a. Storage transfer samples (s t ||g,a t , r t ,s t+1 ||g,p,D traj ) to the experience pool; b. Sample trajectories from the experience pool according to the priority sampling probability of the trajectory being replayed; c. Then sample from the trajectory (s t , a t ,s t+1 ), sample the virtual target state g′∈{s t+1 , ..., s T-1 }; d. Recalculate the reward for the transferred sample; e. Storage of post-transfer samples (s t |g′,a t , r t ′,s t+1 |g′,p,D traj ) to the experience pool D; (6) Sample from the experience pool D to train the network and calculate the q value of the training network and the target network Value, and update the training network parameters θ; In the formula, is the target value of the sample extracted from the experience replay pool, μ and θ are the learning rate and training network parameters respectively; Q(·) and are the state-action functions of the training network and the target network respectively; (7) Copy the training network parameters θ at a fixed frequency to update the target network parameters (8) If t ≥ T, continue with the following steps, otherwise t = t + 1, return to step (3); (9) If e≥E, exit the loop; otherwise, e=e+1 and return to step (3).

7. The method for parallel control of power transmission line construction equipment based on Bi-LSTM and DDPG algorithm according to claim 1, characterized in that: The improved DDPG algorithm is used to construct a decision control model, including: State space: the state of the transmission line construction equipment, including the predicted values ​​of current, voltage and load; the predicted values ​​of the transmission line construction equipment parameters and the equipment operability constraints; Action Space: A t ={δ t ,ΔP t }; In the formula, δ t is the switching action of the circuit breaker and disconnector; ΔP t is the change in power of construction equipment; Reward function: In the formula, w i and w j are the weights of the indicator reward item and penalty item respectively; r i and p j are the i-th reward item value and the j-th penalty item value respectively.

8. A parallel control device for power transmission line construction equipment based on Bi-LSTM and DDPG algorithm, characterized in that: The device is applied to the method described in any one of claims 1 to 7, and the device comprises: The construction equipment status prediction result acquisition module is used to optimize the weight of the temporal attention mechanism based on the differential evolution algorithm of the stage mutation strategy, and construct a Bi-LSTM network model including an input layer, a feature attention layer, a Bi-LSTM network layer, a temporal attention layer, a fully connected layer and an output layer through the feature attention mechanism and the optimized temporal attention mechanism; the collected construction equipment status data is input into the Bi-LSTM network model to obtain the construction equipment status prediction result; The optimal control strategy acquisition module is used to improve the DDPG algorithm based on the post-experience playback of priority, build a decision control model using the improved DDPG algorithm, and input the construction equipment state prediction results into the decision control model to obtain the optimal control strategy; The construction equipment control module is used to control the transmission line construction equipment through the optimal control strategy.

9. A parallel control device for power transmission line construction equipment based on Bi-LSTM and DDPG algorithm, characterized in that: including memory and processor; The memory is used to store computer program code and transmit the computer program code to the processor; The processor is configured to execute the method according to any one of claims 1 to 7 according to instructions in the computer program code.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images