Mine power supply intelligent monitoring platform based on multi-system fusion

Abstract

The invention relates to the technical field of mine power supply monitoring, and discloses a mine power supply intelligent monitoring platform based on multi-system fusion, and the platform comprises the steps: obtaining multi-time-scale operation data, and classifying the data into a fast variable data stream, a medium-speed variable data stream and a slow variable data stream; a three-layer reinforcement learning framework is constructed, the top layer processes minute-level economic dispatching, the middle layer processes second-level path optimization, and the bottom layer executes millisecond-level protection coordination; interlayer coordination is realized through a two-way information transmission mechanism; and dynamically adjusting control parameters by using a rolling horizon optimization algorithm. The invention also discloses a mine power supply intelligent monitoring system based on multi-system fusion. The mine power supply intelligent monitoring system comprises a data acquisition and classification module, a three-layer decision module, a coordination communication module and a dynamic optimization module.

Description

Technical Field

[0001] This invention relates to the field of mine power supply monitoring technology, and more specifically, to a mine power supply intelligent monitoring platform based on multi-system integration. Background Technology

[0002] As a critical infrastructure for safe coal mine production, the operation of the mine power supply system involves complex decision-making processes across multiple time scales. In actual operation, the system needs to simultaneously handle millisecond-level relay protection action coordination, second-level power supply path switching, and minute-level economic operation optimization. These decision-making tasks at different time scales are both independent and interdependent, and each scale has strict real-time requirements.

[0003] Existing technologies typically handle problems at different time scales independently, employing a hierarchical approach to design protection devices, path optimization algorithms, and economic scheduling strategies separately, with a lack of effective information exchange and coordination mechanisms between the different levels.

[0004] However, this independent processing approach ignores the coupling effect across scales. The action strategy of the protection device may conflict with the power supply path optimization results, leading to frequent system switching or protection mismatch. If a unified modeling method is used, it will face the curse of dimensionality problem and cannot meet the millisecond-level response requirements. Therefore, how to achieve coordinated optimization of decisions at different time scales has become an urgent technical problem to be solved in mine power supply monitoring systems. Summary of the Invention

[0005] This invention provides a mine power supply intelligent monitoring method based on multi-system integration, which solves the technical problem of system operation conflict caused by lack of coordination in decision-making at different time scales in related technologies.

[0006] This invention discloses a method for intelligent monitoring of mine power supply based on multi-system fusion, comprising the following steps: acquiring multi-timescale operational data of the mine power supply system; classifying the operational data into fast variable data streams, medium-speed variable data streams, and slow variable data streams according to the data update frequency; constructing a three-layer reinforcement learning architecture including a top-level agent, a middle-level agent, and a bottom-level agent; inputting the slow variable data stream into the top-level agent for minute-level economic scheduling strategy learning and outputting long-term operational constraints; inputting the medium-speed variable data stream into the middle-level agent for second-level path optimization decision-making and generating topology constraints; and inputting the fast variable data stream into the bottom-level agent... The layered intelligent agent executes millisecond-level protection coordination control; inter-layer coordination is carried out based on a two-way information transmission mechanism, using the upper-layer decision results as constraints for the lower-layer constraints to be transmitted from top to bottom, and feeding back the lower-layer execution results to update the upper-layer value estimates to generate time-consistent optimized control sequences; a rolling time-domain optimization algorithm is used to dynamically adjust the prediction time window and optimization step size according to the system state change rate, expanding the prediction time domain when the system is in steady state and shortening the prediction time domain when a disturbance is detected, outputting multi-scale cooperative control commands that include protection setting adjustments, switching operation sequences, and power allocation schemes.

[0007] This invention discloses a mine power supply intelligent monitoring system based on multi-system fusion, used to execute the aforementioned mine power supply intelligent monitoring method based on multi-system fusion. The system includes: a data acquisition and classification module for acquiring multi-timescale operational data of the mine power supply system and classifying it according to update frequency; a three-layer decision-making module comprising top-level, middle-level, and bottom-level agents for handling optimization tasks at different timescales; a coordination and communication module for realizing bidirectional information transmission and constraint coordination between agents; and a dynamic optimization module for adjusting control parameters according to the system state change rate and outputting collaborative control commands.

[0008] Furthermore, the classification of operational data based on data update frequency includes: when the data update cycle is less than 100 milliseconds, it is classified as fast variable data stream, including current and voltage sampling values ​​of protection devices and fault detection signals; when the data update cycle is between 100 milliseconds and 10 seconds, it is classified as medium-speed variable data stream, including switch status changes and network topology switching information; when the data update cycle is greater than 10 seconds, it is classified as slow variable data stream, including load forecast values, electricity price data, and long-term operational statistics.

[0009] Furthermore, the implementation of the three-layer reinforcement learning architecture includes: processing the classified data stream using a wavelet multi-resolution analysis algorithm to separate the state features of different frequency components, generating a hierarchical feature vector group containing fast time-scale feature vectors, medium time-scale feature vectors, and slow time-scale feature vectors; inputting the slow time-scale feature vectors into the top-level policy network to output an economic dispatch action vector and an approximate value of the value function, where the economic dispatch action vector represents the output allocation ratio of each generator unit, and generating the required operating cost constraints and power balance constraints for the middle layer based on this action vector; decomposing the middle-level path optimization problem into multiple subgraphs using a graph decomposition algorithm, inputting the top-level constraints as boundary conditions into a parallel mixed-integer programming solver to generate a set of candidate paths that satisfy the constraints; transforming the middle-level path decision into feasible domain constraints of the bottom-level protection parameters based on a constraint propagation algorithm, inputting them into the bottom-level Q-network to learn the protection action strategy, and the bottom-level Q-network outputs Q-value estimates for three discrete actions: no action, tripping, and issuing an alarm.

[0010] Furthermore, the training of the top-level policy network includes: adopting the Actor-Critic training mode and optimizing policy parameters through the policy gradient algorithm; the policy loss is calculated as the negative expected value of the product of the action advantage function and the logarithm of the policy probability, where the advantage function represents the superiority or inferiority of a certain action relative to the average action in that state; the value loss is in the form of the square of the temporal difference error, and the value estimate is updated by minimizing the mean square error between the value network output and the target value; the total loss function is the weighted sum of the policy loss and the value loss, and the parameters are updated using the Adam optimization algorithm.

[0011] Furthermore, the mid-level path optimization includes: identifying weak connections in the network by calculating the edge betweenness number; determining a weak connection edge as one whose edge betweenness number is less than 0.5 times the average betweenness number of all edges in the network; identifying key split points by the increase in the number of connected components after removing nodes; decomposing the power supply network graph into multiple subgraphs based on weak connections and key split points; constructing a hybrid integer programming model, where integer variables represent the on / off states of switching devices, and continuous variables represent line transmission power and node voltage; and constraints including power flow balance constraints, line capacity constraints, and voltage upper and lower limit constraints. The power flow balance constraint requires that the injected power of each node be equal to zero, the line capacity constraint requires that the line power not exceed 90% of the rated capacity, and the voltage constraint requires that the node voltage be between 95% and 105% of the rated voltage.

[0012] Furthermore, the underlying protection includes: a constraint propagation algorithm determines the currently active set of lines based on candidate power supply paths and identifies the cascading relationship between upstream and downstream protection devices for each line; according to the protection coordination principle, the downstream protection device is required to operate earlier than the upstream protection device, with an operating time difference of not less than 0.3 seconds; the operating current setting of the protection device is greater than 1.2 times the maximum load current and less than 0.8 times the minimum short-circuit current; the underlying Q network adopts an empirical replay and target network training mode, optimizes Q value estimation through temporal difference learning, and the loss function is the expected square of the difference between the current Q value and the target Q value.

[0013] Furthermore, the bidirectional information transmission mechanism includes: in the top-down constraint transmission process, the top layer periodically sends updated economic scheduling constraint parameters to the middle layer, and the middle layer periodically sends updated topology constraint parameters to the bottom layer; in the bottom-up feedback update process, the bottom layer immediately feeds back the action result and reward value to the middle and top layers after each protection action is executed; the middle layer periodically updates the Q-value estimate of the path optimization strategy based on the feedback from the bottom layer, and the top layer periodically updates the value function estimate based on the cumulative feedback from the middle and bottom layers; the update of the value function is achieved through the time-series difference method, and the feedback reward value is accumulated into the value estimate of the corresponding state.

[0014] Furthermore, the rolling time-domain optimization includes: measuring the rate of change of the system state by calculating the time derivatives of key state variables, including node voltage, line current, and load power; determining that the system is in steady state when the absolute value of the rate of change of all key state variables is less than a set threshold, and determining that a disturbance is detected when any rate of change exceeds the threshold; dynamically adjusting the prediction time window according to the system state, with a prediction time domain of 15 minutes and an optimization step size of 60 seconds in steady state, and shortening the prediction time domain to 5 minutes and the optimization step size to 15 seconds during disturbances; solving for the optimal control sequence within the prediction time domain, with the objective function being the weighted sum of the operating costs at each moment in the prediction time domain, the weight coefficients reflecting the time discount effect, and only executing the control action at the current moment.

[0015] Furthermore, the solution for the optimal control sequence includes: transforming the rolling time-domain optimization problem into a quadratic programming problem, with the objective function minimizing the cumulative sum of weighted operating costs within the prediction time domain; operating costs include generation costs, power purchase costs, and switching costs, where generation costs are the sum of the products of the output of each generator unit and its unit cost, power purchase costs are the product of the purchased power and the electricity price, and switching costs are the product of the number of switching actions and the cost of a single switching operation; the weight coefficients decay exponentially with time, decreasing as time progresses, and the discount factor is 0.95; constraints include the system dynamic equations and the constraint parameters passed from each layer, and the optimal control sequence is obtained by using a sequential quadratic programming algorithm.

[0016] This invention utilizes a time-scale-aware three-layer reinforcement learning architecture to decompose complex multi-objective optimization problems into sub-problems with different time granularities. This overcomes the curse of dimensionality faced by traditional unified modeling methods and solves the technical problem of system operational conflicts caused by a lack of coordination in multi-timescale decisions. It achieves the technical effect of enabling millisecond-level protection coordination, second-level path optimization, and minute-level economic scheduling to operate in a coordinated manner. Specifically, through data classification and hierarchical processing, direct solutions to high-dimensional state spaces are avoided; a two-way information transmission mechanism ensures the coordination and consistency between decisions at different timescales, eliminating decision conflicts caused by cross-scale coupling effects; and a rolling time-domain optimization algorithm enables adaptive control of the system under disturbance and steady-state conditions, improving the operating efficiency and reliability of the mine power supply system. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of the intelligent monitoring method for mine power supply based on multi-system integration of the present invention; Figure 2 This is a flowchart illustrating the implementation steps of the three-layer reinforcement learning architecture in this invention; Figure 3 This diagram illustrates the classification of different data types in a mine power supply system and their update cycle distribution. The bar chart represents the number of feature dimensions for each data type, and the line chart represents the average update cycle. Figure 4 The three-layer reinforcement learning architecture was demonstrated to optimize performance in terms of operating cost, power reliability, and response time. The downward trend of the curves indicates that the system continuously improves its performance through learning. Figure 5 The comparison of generator output allocation before and after the top-level intelligent agent's economic scheduling optimization is shown. Figure 6 The dynamic changes in the utilization rate of power supply lines in the three mining areas are shown. In particular, it demonstrates that after the load of mining area 2 suddenly increased at 10:05, the mid-level intelligent agent switched the single-line power supply to three-line parallel connection through path optimization, which reduced the line utilization rate from 128% (overload) to 61% (normal). Detailed Implementation

[0018] As a critical infrastructure for safe coal mine production, the operation of the mine power supply system involves complex decision-making processes across multiple time scales. In actual operation, the system needs to simultaneously handle millisecond-level relay protection action coordination, second-level power supply path switching, and minute-level economic operation optimization. These decision-making tasks at different time scales are both independent and interdependent, and each scale has strict real-time requirements.

[0019] Existing technologies typically treat problems at different time scales independently, neglecting the coupling effects between scales. For example, the action strategy of protection devices may conflict with the power supply path optimization results, leading to frequent system switching or protection mismatch. If a unified modeling method is used, it will face the curse of dimensionality, failing to meet the millisecond-level response requirements. At the same time, the coordination mismatch between multi-level protection devices and the NP-hard nature of power supply path optimization further increase the complexity of the problem.

[0020] like Figure 1-6 As shown, the method of this embodiment includes the following steps: Step 1: Acquire multi-timescale operational data of the mine power supply system and generate categorized data streams. This step acquires system operation data, including protection device action records, network topology status, and load forecast data. By extracting the time stamp of each data point, it is categorized into fast variable data streams (milliseconds), medium-speed variable data streams (seconds), and slow variable data streams (minutes) based on the data update frequency.

[0021] Furthermore, the aforementioned data classification is determined based on the update frequency. Data updates less than 100 milliseconds are classified as fast variable data streams, mainly including current and voltage sampling values ​​of protection devices and fault detection signals. Data updates between 100 milliseconds and 10 seconds are classified as medium-speed variable data streams, mainly including switch status changes and network topology switching information. Data updates greater than 10 seconds are classified as slow variable data streams, mainly including load forecasts, electricity price data, and long-term operation statistics.

[0022] The power supply system of a certain coal mine includes a main substation, three mining area substations, two generator sets, and 15 10kV power supply lines. The system starts data acquisition at 10:00, and the monitoring scope covers the electrical parameters, equipment status, and load information of the entire network.

[0023]

[0024] After data classification, the above data is divided into three data streams according to the update cycle: fast variables such as current and voltage with an update cycle of 20-50ms are classified into the fast variable data stream, forming a 72-dimensional feature vector (including three-phase current of 15 lines, 6 bus voltages, and 15 fault detection signals); switch status and topology identifiers with an update cycle of 2-5 seconds are classified into the medium-speed variable data stream, forming a 28-dimensional feature vector (including 22 switch statuses and 6 topology identifiers); and load forecasts and electricity price information with an update cycle of 60 seconds are classified into the slow variable data stream, forming an 18-dimensional feature vector (including 3 regional loads, 2 generation costs, electricity prices, and 12 statistical parameters).

[0025] Step 2: Based on a three-layer reinforcement learning architecture, output a multi-scale optimization control strategy. This step constructs a three-layer reinforcement learning architecture comprising a top-level agent, a middle-level agent, and a bottom-level agent. The top-level agent receives slow-variable data streams and learns economic scheduling strategies at the minute level; the middle-level agent receives medium-speed variable data streams and processes path optimization decisions at the second level; the bottom-level agent receives fast-variable data streams and executes protection and coordination control at the millisecond level.

[0026] It should be noted that, in addition to step 2, the following steps are also included: Step 201: Use wavelet multi-resolution analysis algorithm to process the classification data stream, separate the state features of different frequency components, and generate hierarchical feature vector groups; The aforementioned wavelet multi-resolution analysis algorithm takes fast variable data stream, medium-speed variable data stream and slow variable data stream as input, performs wavelet decomposition on each data stream, decomposes the signal into approximate components and detail components at different frequency scales, extracts the feature coefficients at each scale, and outputs a hierarchical feature vector group containing fast time scale feature vectors, medium time scale feature vectors and slow time scale feature vectors.

[0027] Furthermore, the aforementioned wavelet decomposition uses the Daubechies wavelet basis function, with a decomposition layer of 3. The first layer extracts high-frequency detail components corresponding to fast time-scale features, the second layer extracts mid-frequency components corresponding to mid-time-scale features, and the third layer extracts approximate components corresponding to slow time-scale features.

[0028] Step 202: Input the state features of the slow time scale into the top-level policy network, and output the long-term operating constraints and approximate value of the value function; The aforementioned top-level strategy network employs a multi-layer fully connected neural network. Its input layer receives a state feature vector with a slow time scale. The state feature vector contains load forecasts, electricity price information, and generation cost parameters related to economic dispatch. Since these parameters have different dimensions and numerical ranges, mean normalization based on the range is performed on each parameter before it is input into the network, scaling all input features to the same numerical range.

[0029] Furthermore, the dimension of the aforementioned state feature vector is equal to the number of monitoring parameters, specifically including: the future load forecast value of each power supply area, the electricity price of the current period, the unit power generation cost of each generator set, the current output status of each generator set, and the available capacity; the value range of each element of the state vector is the physical feasible range of that parameter, for example, the value range of the load forecast value is from zero to the maximum load demand of the area, and the value range of the generator set output value is from zero to the rated capacity of the unit.

[0030] The top-level policy network contains multiple hidden layers, each of which undergoes a nonlinear transformation using the ReLU activation function. The output layer is divided into two branches. The first branch outputs the economic dispatch action, which is a continuous vector representing the output allocation ratio of each generator unit. The network output is converted into the actual power value through inverse normalization mapping, and the operating cost constraints and power balance constraints required by the middle layer are generated based on the power allocation scheme. The second branch outputs an approximate value of the value function.

[0031] Furthermore, the dimension of the aforementioned economic dispatch action vector is equal to the number of generator sets. Each element represents the output allocation ratio of the corresponding generator set, with a value range of [0,1], representing the output ratio of the unit from zero to full load. After inverse normalization, the m-th element of the action vector is converted into the actual output power. ,in Indicates the generator set number. For normalized action values, Let m be the rated capacity of the m-th generator set.

[0032] Furthermore, the aforementioned long-term operational constraints are derived from economic dispatch actions, specifically: the total operating cost is calculated based on the output allocation of the generator sets, and the total operating cost is used as the upper limit constraint of the cost for mid-level path optimization; the power margin of each node is calculated based on the output power of the generator sets and the load forecast value, and the power margin is used as the constraint range of the line transmission power, thereby providing the mid-level with dual constraints of economy and reliability.

[0033] The top-level policy network uses the Actor-Critic training mode, optimizes policy parameters through the policy gradient algorithm, and the loss function consists of policy loss and value loss. The policy loss is calculated as follows:

[0034] in, For the dominant function, This represents the action probability distribution output by the policy network.

[0035] Furthermore, the aforementioned expectation operation represents the average of all state-action sample pairs in the training batch. During training, multiple historical state-action pairs are sampled from the experience replay buffer, and the loss of each sample is calculated and then averaged as the batch loss.

[0036] Furthermore, the aforementioned advantage function is calculated as follows: ,in For state-action value function, The state-action value function is the state-value function, and the dominance function measures the superiority or inferiority of a particular action relative to the average action in that state. The aforementioned state-action value function is estimated using a temporal difference method, specifically as follows: The immediate reward obtained after the current action is performed is Next state The value is output through the value network.

[0037] Furthermore, the aforementioned top-level intelligent agent's immediate reward is calculated based on the economic scheduling effect, specifically the amount of saving in operating cost relative to the baseline cost at the current moment. Operating cost includes generation cost and electricity purchase cost. Generation cost is the sum of the product of the output of each generator unit and its unit cost, while electricity purchase cost is the product of the power purchased from the grid and the electricity price. The baseline cost is the cost of operating according to the fixed scheduling strategy. When the saving amount is positive, the reward is positive, indicating that the scheduling strategy optimization is effective. When the saving amount is negative, the reward is negative, indicating that the scheduling strategy needs to be improved.

[0038] The value loss is expressed in the form of the square of the time-series difference error:

[0039] in, For value network output, For the target value, calculate as , For instant rewards, This is the discount factor, with a value range of [0.95, 0.99]. In this embodiment, the value is 0.98. The next state is defined as follows: the total loss function is the weighted sum of the policy loss and the value loss, with a weighting coefficient of 0.5, and the Adam optimization algorithm is used for parameter updates.

[0040] Step 203: Use the graph decomposition algorithm to decompose the mid-level path optimization problem into multiple subgraphs, and input the top-level constraints as boundary conditions into the parallel mixed-integer programming solver to generate a candidate path set; The aforementioned graph decomposition algorithm takes the topology and node connection relationships of the mine power supply network as input, identifies weak connection edges and key split points in the network, and decomposes the power supply network graph into multiple relatively independent subgraphs according to the topology. Each subgraph contains nodes and edges of a local power supply area, and outputs multiple smaller subgraph structures.

[0041] Furthermore, the aforementioned weak connection edge identification is achieved by calculating the edge betweenness number, which is defined as the number of times the shortest path between all pairs of nodes in the network passes through that edge. The Floyd-Warshall algorithm is used to calculate the shortest path between each pair of nodes, and the number of times each edge is passed by the shortest path is used as the edge betweenness number. When the edge betweenness number is less than 0.5 times the average betweenness number of all edges in the network, it is determined to be a weak connection edge. Key split points are identified by the increase in the number of connected components after removing a node. If the number of connected components increases after removing a node, then that node is a key split point.

[0042] The aforementioned mixed-integer programming solver takes the topology information of each subgraph, the top-level economic scheduling constraints, and the power supply reliability constraints as inputs, constructs an optimization model containing integer and continuous variables, solves the power supply paths of each subgraph in parallel, and outputs a set of candidate paths that meet the constraints.

[0043] Furthermore, in the aforementioned mixed-integer programming model, integer variables represent the on / off states of switching equipment, taking values ​​of 0 or 1, while continuous variables represent line transmission power and node voltage. Constraints include power flow balance constraints, line capacity constraints, and voltage upper and lower limit constraints.

[0044] Furthermore, the aforementioned power flow balance constraint requires that the injected power of each node be equal to zero. For node i, the following condition must be met: ,in Indicates the node number. This represents the node connected to node i. Let i be the set of all nodes connected to node i. For the line power flowing from node i to node j, when When it is an upstream node A negative value indicates an inflow. Let i be the power generation of node i. Let be the load power of node i; the line capacity constraint is expressed as: ,in The rated capacity of line ij is given; the voltage constraint is expressed as... ,in Let be the voltage at node i. This is the rated voltage.

[0045] Step 204: Based on the constraint propagation algorithm, the mid-level path decision is transformed into the feasible domain constraint of the bottom-level protection parameters, and then input into the bottom-level Q network to learn the protection action strategy.

[0046] The aforementioned constraint propagation algorithm takes the candidate power supply path and network topology switching scheme output by the mid-level optimization as input, analyzes the impact of path switching on the action timing and settings of protection devices, derives the parameter constraint range of protection devices at all levels according to the protection coordination principle, transforms the topology constraints implicit in the upper-level decision into the upper and lower bound constraints of the bottom-level protection parameters, and outputs the feasible domain constraints of the bottom-level protection parameters.

[0047] Furthermore, the aforementioned constraint propagation process first determines the set of currently active lines based on candidate power supply paths. For each active line, the cascading relationship between its upstream and downstream protection devices is identified. Then, according to the protection coordination principle, the action time of the downstream protection device should be earlier than that of the upstream protection device, and the action time difference is set to be no less than 0.3 seconds. From this, the upper and lower bounds of the action time of each level of protection device are derived. Finally, the range of the operating current setting of the protection device is determined based on the load prediction value and the short-circuit current calculation value of the line. These time and current constraint ranges are used as feasible domain constraints for output.

[0048] Furthermore, the aforementioned maximum load current of the line is obtained through statistical analysis of historical load data, and the 95th percentile value of the load current of the line in the historical data is taken as the maximum load current; the aforementioned minimum short-circuit current is obtained through short-circuit calculation, and the short-circuit current when a two-phase short circuit occurs at the end of the line is calculated using the nodal admittance matrix method as the minimum short-circuit current; the operating current setting value of the protection device is required to be greater than 1.2 times the maximum load current and less than 0.8 times the minimum short-circuit current.

[0049] The aforementioned underlying Q-network adopts a deep Q-learning architecture. Its input layer receives a state vector, which contains the current measurement value of the protection device, network topology information, and feasible domain constraint parameters passed by the middle layer. Since the measurement values ​​contain physical quantities such as current and voltage with different dimensions, the various measurement values ​​are standardized before being input into the network. At the same time, the network topology information is converted into numerical features through one-hot encoding.

[0050] Furthermore, the aforementioned bottom-level state vector consists of: the amplitude and phase angle of the three-phase current monitored by the protection device, the amplitude and phase angle of the bus voltage, the fault detection flag, the one-hot encoded vector of the current network topology, the upper and lower bounds of the action time constraint and the upper and lower bounds of the action current setting constraint transmitted by the middle layer; the dimension of the state vector is equal to the sum of the number of elements of each of the above components, the range of current and voltage measurements is from zero to the upper limit of the sensor's range, the range of phase angle is [0, 2π], and the range of constraint parameters is determined by the calculation results of the middle layer.

[0051] The underlying Q-network contains multiple fully connected hidden layers and uses the ReLU activation function for nonlinear mapping. The output layer corresponds to a discrete protection action space, outputting the Q-value estimate for each action. Based on the ε-greedy strategy, the action with the largest Q-value is selected as the actual protection action to be executed.

[0052] Furthermore, the aforementioned protection action space includes three discrete actions: no action, tripping, and issuing an alarm. The output layer of the Q network contains three neurons, each corresponding to a Q value for one of these three actions. The Q value ranges from negative infinity to positive infinity, representing the long-term cumulative reward estimate for performing the action. The ε-greedy strategy explores actions randomly with a low probability and selects the action with the largest Q value with a high probability. As training progresses, the exploration probability is gradually reduced to converge to the optimal strategy.

[0053] The underlying Q-network employs an empirical replay and target network training mode, optimizing Q-value estimation through temporal difference learning. The loss function is defined as:

[0054] in, This represents the Q-value output by the current network. For instant rewards, This is a discount factor with a value of 0.99. Indicates the optional action for the next state. The maximum Q-value of the next state output by the target network.

[0055] Furthermore, the aforementioned immediate reward is calculated based on the correctness of the protection action and the response time. The correctness of the protection action is determined by matching the actual action with a preset action rule library. The action rule library predefines the correct actions for various scenarios based on the fault type and fault location. When the executed action is consistent with the correct action in the rule library and the response time does not exceed 50 milliseconds, the reward is positive. When the action is inconsistent or the response times out, the reward is negative.

[0056] By minimizing the loss function to make the Q value approximate the solution of the Bellman optimal equation, the Adam optimization algorithm is used to update the network parameters; the target network parameters are periodically copied from the current network to stabilize the training process.

[0057] In this embodiment of the application, in order to improve the learning efficiency of each layer of intelligent agents, step 205 is also included: updating the policy network parameters of each layer online using the temporal difference learning algorithm, and storing cross-scale interaction data through the experience replay buffer.

[0058] The aforementioned temporal difference learning algorithm uses the state transition sequence generated by the interaction between the agent and the environment as input to calculate the temporal difference error. ,in For instant rewards, For the value estimation of the next state, The value of the current state is estimated, and the network parameters are updated using the gradient descent method based on this error, and the updated network parameters are output. The aforementioned experience replay buffer stores the state, action, reward and interaction experience of each layer of the agent in the next state, and random samples of historical experience are used for training.

[0059] At 10:02, the system detected that the load of mining area 3 would increase from the current 1.2MW to 2.5MW within the next 15 minutes, requiring the initiation of multi-scale optimization decision-making. The three-layer architecture processes the corresponding data in the classified data streams respectively.

[0060] The top-level agent receives a slow variable data stream. The state vector contains the load forecast for the next 15 minutes, the current electricity price, and the generation cost. The slow timescale feature vector extracted after wavelet decomposition is input into the top-level policy network. After normalization, the state features are as follows: load forecast for mining area 1 (1.45MW) is normalized to 0.48; load forecast for mining area 2 (1.85MW) is normalized to 0.62; load forecast for mining area 3 (2.50MW) is normalized to 0.83; electricity price (0.65 yuan / kWh) is normalized to 0.54; cost of generator No. 1 (0.45 yuan / kWh) is normalized to 0.38; and cost of generator No. 2 (0.52 yuan / kWh) is normalized to 0.43.

[0061] Table 2. Economic Scheduling Decision Output of the Top-Level Agent

[0062] The economic dispatch scheme output by the top-level strategy network increases the output of Unit 1 from 1.5MW to 2.8MW and decreases the output of Unit 2 from 2.0MW to 1.9MW, reducing grid power purchases from 1.35MW to 1.1MW. This scheme reduces operating costs from 3.12 yuan per second to 2.75 yuan per second, achieving a savings rate of 11.9%, while simultaneously generating a constraint that the total operating cost does not exceed 2.85 yuan per second and passing it to the middle layer. The value function branch of the top-level strategy network outputs an approximate value of 125.3 for the current state, representing the expected long-term cumulative reward for executing the optimal strategy in the current state.

[0063] The middle-level agent receives the medium-speed variable data stream and the constraints passed from the top level, and processes the power supply path optimization. The current network topology is topology A, which includes the main power supply line and two backup lines. The graph decomposition algorithm divides the power supply network into three subgraphs, corresponding to the power supply areas of the three mining zones. The mixed-integer programming solver solves for the optimal power supply path in each subgraph while satisfying the top-level cost constraints.

[0064] Table 3 Path optimization results for mid-level agents

[0065] After the mid-level optimization, mining area 2 and mining area 3 were changed from single-line power supply to dual-line parallel power supply to meet the load growth demand, while the path switching scheme and topology constraints were passed to the bottom layer.

[0066] The bottom-level agent receives fast variable data streams and topology constraints passed from the middle layer, and performs protection coordination control. The constraint propagation algorithm calculates the feasible region of protection parameters based on the new power supply path. The maximum load current of line L6 is 2100A, and the minimum short-circuit current is 8500A. From this, the protection action current setting range is derived to be 2520-6800A, and the upper and lower bounds of the action time are 0.05-0.35 seconds. These feasible region constraints, together with the current electrical quantity state, constitute the input state vector of the bottom-level Q network.

[0067] During the path switching preparation phase, the underlying Q network continuously monitors the electrical status of each line. At 10:02:05, after the state vectors of lines L6 and L7 are input into the Q network, the network calculates the Q-value estimates for the three actions of the protection devices for each line.

[0068] Table 4. Protection Action Decisions of the Bottom Agent

[0069] The underlying Q-network selects the action with the largest Q value as the actual protection action based on the ε-greedy strategy. Since the electrical quantities of all lines are currently within the normal range and the fault detection flag is normal, the Q value of "no action" (8.28-8.52) is significantly higher than that of "trip" (-11.5 to -12.3) and "alarm" (3.12-3.21). Therefore, all protection devices output "no action" commands to maintain stable system operation.

[0070] Step 3: Perform inter-layer coordination based on a two-way information transmission mechanism to generate time-consistent optimized control sequences. In this step, the decision-making results of the upper layer serve as constraints for the lower layer, and the execution results of the lower layer are fed back to update the value estimate of the upper layer. The economic scheduling strategy of the top layer provides operating cost constraints for the middle layer, the path optimization results of the middle layer provide topological constraints for the lower layer, and the execution status of the protection actions of the lower layer is fed back to the middle and top layers, forming a closed-loop two-way information transmission.

[0071] Furthermore, the aforementioned two-way information transmission mechanism includes two processes: top-down constraint transmission and bottom-up feedback update. During constraint transmission, the top layer periodically sends updated economic scheduling constraint parameters to the middle layer, and the middle layer periodically sends updated topology constraint parameters to the bottom layer. During feedback update, the bottom layer immediately feeds back the action result and reward value to the middle and top layers after each protection action is executed. The middle layer updates the path optimization strategy periodically based on the feedback from the bottom layer, and the top layer updates the value function estimate periodically based on the cumulative feedback from the middle and bottom layers. The update of the value function is achieved by accumulating the feedback reward value into the value estimate of the corresponding state.

[0072] At 10:02:15, the bottom-level protection device performs a path switching operation, and switches K12-K15 close sequentially according to the instructions from the middle level, realizing the parallel power supply of dual lines to mining area 3. The bottom-level intelligent agent feeds back the execution result to the upper level.

[0073] During the bottom-up feedback update process, after the bottom-level protection device completed the path switching, it monitored that the currents of lines L6 and L7 stabilized at 1280A and 1220A respectively, indicating a basically balanced power distribution between the two lines. There were no overcurrent or voltage fluctuations during the switching process, and the protection action was executed correctly. The bottom-level device calculated an immediate reward of +15.3, which was fed back to the middle and top levels.

[0074] Table 5 Data on Inter-layer Feedback Information Transmission

[0075] After receiving feedback from the lower layer, the middle layer updates the Q-value estimate of the path optimization strategy. The Q-value of the dual-line parallel power supply scheme in mining area 3 increases from 7.85 to 8.12, indicating an improvement in the long-term value assessment of this path strategy. Simultaneously, the middle layer feeds back the path optimization completion information and the accumulated reward of +22.7 to the top layer.

[0076] After receiving feedback from the middle and bottom layers, the top layer updates the value function estimate. The value function for the current state "load increase + electricity price of 0.65 yuan / kWh" is updated from 125.3 to 148.0. This update is achieved through the time-series difference method and is calculated as follows: The learning rate Timing difference error Therefore, the updated value is .

[0077] During the top-down constraint propagation process, the top layer re-evaluates the economic scheduling strategy based on the updated value function and sends new constraint parameters to the middle layer at 10:03:00. Due to the continuous increase in load, the top layer adjusts the upper limit of the operating cost constraint from 2.85 yuan / s to 3.10 yuan / s, allowing the middle layer to optimize the path within a wider cost range.

[0078] Table 6 Constraint Transfer Process Data

[0079] After receiving the new constraints from the top layer, the middle layer expands the path search space and resolves the mixed-integer programming model. Under relaxed cost constraints, the middle layer maintains the current path scheme because the existing scheme already satisfies the new constraints and operates stably. The middle layer transmits topological constraints to the bottom layer, requiring that switches K12-K15 remain closed, and the bottom layer maintains its current protection settings accordingly.

[0080] Through bidirectional information transmission, the three-layer intelligent agent forms a coordinated control strategy, ultimately generating a time-consistent optimized control sequence at 10:03:05: the top layer maintains the economic dispatch scheme of Unit 1 outputting 2.8MW and Unit 2 outputting 1.9MW; the middle layer maintains the path configuration of single-line power supply for mining area 1 and dual-line parallel power supply for mining areas 2 and 3; the bottom layer keeps all protection devices in a "non-operation" state and maintains their current protection settings. This control sequence maintains consistency across three time scales: minute-level (economic dispatch), second-level (path configuration), and millisecond-level (protection coordination). The long-term economic goals of the top layer influence the decisions of the middle and bottom layers through constraint transmission, and the execution effect of the bottom layer updates and optimizes the upper-level strategy through feedback, achieving cross-time scale collaborative optimization.

[0081] Step 4: Dynamically adjust control parameters using a rolling time-domain optimization algorithm and output multi-scale cooperative control commands. This step dynamically adjusts the prediction time window based on the system state change rate, adaptively switching the optimization granularity. When the system is in steady state, the prediction time domain is expanded to optimize long-term operating costs; when a disturbance is detected, the time domain is shortened for a faster response. The final output includes multi-scale coordinated control commands for protection setting adjustments, switching operation sequences, and power allocation schemes.

[0082] Furthermore, the aforementioned system state change rate is measured by calculating the time derivative of key state variables, which include node voltage, line current, and load power. The state change rate is defined as the difference between the current state value and the previous state value divided by the time interval. When the absolute value of the change rate of all key state variables is less than a set threshold, the system is determined to be in steady state. The set threshold is 0.5% of the rated voltage per second for the voltage change rate, 1% of the rated current per second for the current change rate, and 2% of the rated power per second for the power change rate. When the absolute value of the change rate of any key state variable exceeds the threshold, a disturbance is detected.

[0083] Furthermore, the prediction time window is dynamically adjusted according to the system state. When the system is in a steady state, the prediction time domain and optimization step size are expanded, and when a disturbance is detected, the prediction time domain and optimization step size are shortened. The adjustment of the prediction time domain and optimization step size is achieved through linear interpolation to achieve a smooth transition.

[0084] The aforementioned rolling time-domain optimization algorithm takes the current system state, the predicted load value within the future time window, and the system dynamic model as inputs, solves the optimal control sequence within the prediction time domain, executes only the control action at the current moment, and re-optimizes based on the new system state at the next moment. Dynamic feedback adjustment is achieved by rolling forward the time-domain window, and the optimized control instructions at each moment are output.

[0085] Furthermore, the aforementioned system dynamic model uses state-space equations to describe the dynamic characteristics of the power supply system. The state variables include node voltages and line currents, and the control variables include switch states and generator outputs. The dynamic equations describe the evolution of the state variables with respect to the control variables and time.

[0086] Furthermore, the aforementioned optimal control sequence solution is achieved by transforming the rolling time-domain optimization problem into a quadratic programming problem. The objective function is the sum of the operating costs at each time step within the prediction time domain, specifically expressed as follows: ,in For the current moment, To predict the length of the time domain, This represents the discrete time point in the prediction time domain. For a moment Operating costs For a moment The weighting coefficients; the aforementioned operating costs It includes three parts: generation cost, power purchase cost, and switching cost. The generation cost is per unit time. The sum of the products of the output of each generator unit and its corresponding unit cost, with the electricity purchase cost being the cost per hour. The product of the purchased power capacity and the electricity price, the switching cost is the time. The product of the number of switching actions compared to the previous moment and the cost of a single switching action; the aforementioned weighting coefficient reflects the time discount effect, with the weight decreasing as time progresses further back in time, calculated as follows: ,in The discount factor is 0.95; the constraints are the system dynamic equations and the constraint parameters passed from each layer. The optimal control sequence is obtained by solving the sequential quadratic programming algorithm.

[0087] At 10:05:00, the system detected a sudden increase in the load power of mining area 2 from 1.85MW to 2.35MW, and the corresponding current in line L3 rose from 1520A to 1950A, with a current change rate of... The voltage change rate exceeded the set threshold of 1.0% / s. Simultaneously, the voltage in mining area 1 decreased from 10.45kV to 10.28kV (voltage change rate: -0.283% / s), and the voltage in mining area 2 decreased from 10.42kV to 10.38kV (voltage change rate: -0.064% / s), both within the threshold. Because the current change rate of line L3 exceeded the threshold, the system determined that a disturbance was detected and initiated rolling time-domain optimization adjustment. The rolling time-domain optimization algorithm adjusts the prediction time domain and optimization step size according to the state change, shortening the prediction time domain from 15 minutes in steady state to 5 minutes, and the optimization step size from 60 seconds to 15 seconds to improve response speed.

[0088] Within the prediction time domain T = 5 minutes (20 steps), the system solves for the optimal control sequence. At the current time t = 10:05:00, the system state is "Mining Area 2 load 2.35MW, electricity price 0.65 yuan / kWh". The objective function calculates the cumulative operating cost over the next 20 time steps. Considering the time discount effect, the weighting coefficients decrease with time, and the weight at time step 1 is... Weight at time 10 Weight at time 20 .

[0089] The sequential quadratic programming algorithm solves for the optimal control sequence for the next 5 minutes, but only executes the control command for the current time 10:05:00. The control commands include adjusting the output of generator No. 1 to 3.2MW, switching the power supply path of mining area 2 from "L3 single line" to "L3+L4+L5 three lines in parallel", and adjusting the protection settings of lines L3-L5.

[0090] Table 7 Control sequences of rolling time-domain optimization output (partial time points)

[0091] The system only executes the control command at 10:05:00 and outputs a multi-scale coordinated control command sequence, including the top-level power allocation command "Adjust the output of Unit 1 to 3.2MW", the middle-level path switching command "Close switches K6 and K9 to form a three-line parallel connection", and the bottom-level protection setting adjustment command "Adjust the upstream protection current setting of L3 to 3200A, with an action time of 0.40s".

[0092] At the next time step, 10:05:15, the system re-performed rolling time-domain optimization based on the new state. At this point, the load in mining area 2 had stabilized at 2.35MW, and the currents of lines L3-L5 were 810A, 780A, and 760A, respectively. The parallel power supply of the three lines reduced the current of a single line, and the line utilization rate dropped from 128% (overload) for a single line to an average of 61% (normal) for all three lines. The system state change rate fell back below the threshold, and the system was re-determined to be in steady state. The prediction time domain was restored to 15 minutes, and the optimization step size was restored to 60 seconds.

[0093] Through dynamic adjustments of rolling time-domain optimization, the system responds quickly when disturbances occur, shortens the prediction time domain and optimization step size to improve decision frequency, and resumes long time-domain optimization after the disturbance is eliminated to take into account economy, thus realizing adaptive switching of optimization granularity under different operating conditions.

[0094] This implementation uses a time-scale-aware three-layer reinforcement learning architecture, which can decompose complex multi-objective optimization problems into sub-problems with different time granularities, overcoming the curse of dimensionality problem faced by traditional unified modeling methods, and thus solving the real-time problem of multi-timescale decision-making.

[0095] Specifically, by classifying system data according to time scales and assigning it to agents at corresponding levels for processing, direct solutions to high-dimensional state spaces are avoided, making millisecond-level protection coordination possible. A two-way information transmission mechanism ensures consistency between decisions at different time scales; the long-term optimization objectives of the upper layer effectively constrain the short-term decisions of the lower layer, while the execution feedback from the lower layer promptly corrects the strategies of the upper layer, thereby eliminating decision conflicts caused by cross-scale coupling effects.

Claims

1. A method for intelligent monitoring of mine power supply based on multi-system fusion, characterized in that, Includes the following steps: Acquire multi-timescale operational data of the mine power supply system, and classify the operational data into fast variable data streams, medium-speed variable data streams, and slow variable data streams according to the data update frequency; A three-layer reinforcement learning architecture consisting of a top-level agent, a middle-level agent, and a bottom-level agent is constructed. The slow variable data stream is input into the top-level agent to learn the economic scheduling strategy at the minute level and output long-term operating constraints. The medium-speed variable data stream is input into the middle-level agent to make path optimization decisions at the second level and generate topology constraints. The fast variable data stream is input into the bottom-level agent to execute millisecond-level protection and coordination control. Inter-layer coordination is carried out based on a two-way information transmission mechanism. The decision-making results of the upper layer are used as constraints of the lower layer for top-down constraint transmission. The execution results of the lower layer are fed back to update the value estimate of the upper layer for bottom-up feedback update, generating a time-consistent optimized control sequence. The rolling time-domain optimization algorithm dynamically adjusts the prediction time window and optimization step size according to the system state change rate. When the system is in steady state, the prediction time domain is expanded, and when a disturbance is detected, the prediction time domain is shortened. The output includes multi-scale coordinated control commands that include protection setting adjustment, switching operation sequence and power allocation scheme.

2. The intelligent monitoring method for mine power supply based on multi-system fusion according to claim 1, characterized in that, The classification of runtime data based on data update frequency includes: Data updates with a cycle of less than 100 milliseconds are classified as fast variable data streams, including current and voltage sampling values ​​of protection devices and fault detection signals. Data updates with a cycle between 100 milliseconds and 10 seconds are classified as medium-speed variable data streams, including switch state changes and network topology switching information. Data updates with a cycle longer than 10 seconds are classified as slow variable data streams, including load forecasts, electricity price data, and long-term operational statistics.

3. The intelligent monitoring method for mine power supply based on multi-system fusion according to claim 1, characterized in that, The implementation of the three-layer reinforcement learning architecture includes: The wavelet multi-resolution analysis algorithm is used to process the classified data stream, separate the state features of different frequency components, and generate a hierarchical feature vector group containing fast time-scale feature vectors, medium time-scale feature vectors and slow time-scale feature vectors. The slow time-scale feature vector is input into the top-level policy network, and the economic scheduling action vector and the approximate value of the value function are output. The economic scheduling action vector represents the output allocation ratio of each generator unit. Based on the action vector, the operating cost constraints and power balance constraints required by the middle layer are generated. The mid-level path optimization problem is decomposed into multiple subgraphs using a graph decomposition algorithm. The top-level constraints are then input as boundary conditions into a parallel mixed-integer programming solver to generate a set of candidate paths that satisfy the constraints. Based on the constraint propagation algorithm, the mid-level path decision is transformed into the feasible domain constraint of the bottom-level protection parameters, which is then input into the bottom-level Q-network to learn the protection action strategy. The bottom-level Q-network outputs the Q-value estimates of three discrete actions: no action, tripping, and issuing an alarm.

4. The intelligent monitoring method for mine power supply based on multi-system fusion according to claim 3, characterized in that, The training of the top-level policy network includes: The Actor-Critic training mode is adopted, and the policy parameters are optimized through the policy gradient algorithm. The strategy loss is calculated as the negative expected value of the product of the action advantage function and the logarithm of the strategy probability, where the advantage function represents the superiority or inferiority of a certain action relative to the average action in that state. The value loss is expressed in the form of the square of the temporal difference error, and the value estimate is updated by minimizing the mean square error between the value network output and the target value. The total loss function is a weighted sum of policy loss and value loss, and the Adam optimization algorithm is used for parameter updates.

5. The intelligent monitoring method for mine power supply based on multi-system fusion according to claim 3, characterized in that, The mid-level path optimization includes: Weak edges in a network are identified by calculating the edge betweenness number. An edge is considered a weak edge when its betweenness number is less than 0.5 times the average betweenness number of all edges in the network. Key split points are identified by the increase in the number of connected components after removing nodes, and the power supply network graph is decomposed into multiple subgraphs based on weak connections and key split points. Construct a hybrid integer programming model, where integer variables represent the on / off states of switching equipment, and continuous variables represent line transmission power and node voltage; The constraints include power flow balance constraints, line capacity constraints, and voltage upper and lower limit constraints. Power flow balance constraints require that the injected power at each node be zero. Line capacity constraints require that the line power not exceed 90% of the rated capacity. Voltage constraints require that the node voltage be between 95% and 105% of the rated voltage.

6. The intelligent monitoring method for mine power supply based on multi-system fusion according to claim 3, characterized in that, The underlying protection includes: The constraint propagation algorithm determines the set of currently active lines based on candidate power supply paths and identifies the cascading relationship between upstream and downstream protection devices of each line; According to the principle of protection coordination, the downstream protection device is required to act earlier than the upstream protection device, and the time difference between the actions should not be less than 0.3 seconds. The operating current setting of the protection device is greater than 1.2 times the maximum load current and less than 0.8 times the minimum short-circuit current; The underlying Q-network adopts an experience replay and target network training mode, and optimizes Q-value estimation through temporal difference learning. The loss function is the expected square of the difference between the current Q-value and the target Q-value.

7. The intelligent monitoring method for mine power supply based on multi-system fusion according to claim 1, characterized in that, The two-way information transmission mechanism includes: In the top-down constraint transmission process, the top layer periodically sends updated economic scheduling constraint parameters to the middle layer, and the middle layer periodically sends updated topology constraint parameters to the bottom layer. In the bottom-up feedback update process, the bottom layer immediately feeds back the action results and reward values ​​to the middle and top layers after each protection action is executed; The middle layer updates the Q-value estimate of the path optimization strategy periodically based on feedback from the bottom layer, and the top layer updates the value function estimate periodically based on the cumulative feedback from the middle and bottom layers. The value function is updated using a time-difference method, which accumulates the feedback reward value into the value estimate of the corresponding state.

8. The intelligent monitoring method for mine power supply based on multi-system fusion according to claim 1, characterized in that, The rolling time-domain optimization includes: The rate of change of system state is measured by calculating the time derivatives of key state variables, which include node voltage, line current and load power. The system is considered to be in steady state when the absolute value of the rate of change of all key state variables is less than the set threshold; when any rate of change exceeds the threshold, a disturbance is detected. The prediction time window is dynamically adjusted according to the system state. In steady state, the prediction time domain is 15 minutes and the optimization step size is 60 seconds. During disturbances, the prediction time domain is shortened to 5 minutes and the optimization step size is shortened to 15 seconds. The optimal control sequence is solved within the prediction time domain. The objective function is the weighted sum of the operating costs at each time point within the prediction time domain. The weight coefficients reflect the time discount effect, and only the control action at the current time point is executed.

9. The intelligent monitoring method for mine power supply based on multi-system fusion according to claim 8, characterized in that, The solution to the optimal control sequence includes: The rolling time-domain optimization problem is transformed into a quadratic programming problem, with the objective function being to minimize the sum of weighted running costs within the prediction time domain; Operating costs include generation costs, electricity purchase costs, and switching costs. Generation costs are the sum of the products of the output of each generator unit and its unit cost. Electricity purchase costs are the product of the purchased power and the electricity price. Switching costs are the product of the number of switching operations and the cost per switching operation. The weighting coefficient decays exponentially over time, with the weight decreasing as time progresses further back in time. The discount factor is 0.

95. The constraints include the system dynamic equations and the constraint parameters passed from each layer. The optimal control sequence is obtained by solving the sequential quadratic programming algorithm.

10. A mine power supply intelligent monitoring platform based on multi-system fusion, used to execute the mine power supply intelligent monitoring method based on multi-system fusion as described in any one of claims 1-9, characterized in that, include: The data acquisition and classification module is used to acquire multi-timescale operational data of the mine power supply system and classify it according to the update frequency. The three-layer decision-making module includes top-level, middle-level, and bottom-level intelligent agents that handle optimization tasks at different time scales; The coordination and communication module is used to realize two-way information transmission and constraint coordination between intelligent agents; The dynamic optimization module is used to adjust control parameters and output coordinated control commands based on the rate of change of system state.

Images