Intelligent operation and maintenance system of box-type substation based on big data

Abstract

This invention provides an intelligent operation and maintenance system for prefabricated substations based on big data, belonging to the field of prefabricated substation operation and maintenance technology. The data access unit unifies the time alignment of electrical operation, environmental monitoring, and historical operation and maintenance data, constructing a constrained time-series data structure based on equipment topology; the state evolution modeling unit extracts multi-scale features, generating three sets of evolution parameters: impedance, capacity, and load; the control variable reconstruction unit generates enhanced control input vectors through component mapping and iterative feedback; the real-time control execution unit outputs reactive power regulation and voltage control commands accordingly; and the human-machine feedback unit embeds manually corrected information and updates control parameters. This invention solves the problems of lack of topological constraints in multi-source data and the difficulty of embedding human experience into automated control, improving the intelligence level, control accuracy, and adaptability of prefabricated substation operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of operation and maintenance technology for prefabricated substations, specifically to an intelligent operation and maintenance system for prefabricated substations based on big data. Background Technology

[0002] Prefabricated substations have tightly coupled internal equipment, enclosed spaces, and are mostly in unattended environments. Their operation and maintenance systems typically collect electrical operation data, environmental monitoring data, and historical operation and maintenance records, and perform reactive power and voltage control based on threshold judgments or fixed logic.

[0003] However, existing technologies still face core challenges in prefabricated substation applications: First, the electrical coupling between equipment nodes is extremely strong, and the lack of a unified processing method that simultaneously reflects topological constraints and dynamic propagation characteristics when using multi-source heterogeneous data for control decisions makes it difficult to accurately describe the transmission path of abnormal states in a compact topology, resulting in delayed or unstable control commands. Second, the field experience of maintenance personnel can only intervene through single command coverage, failing to be continuously embedded in the dynamic generation process of control parameters and adaptively adjusted according to operating conditions to form a complete evolution of control logic. These two problems are coupled, resulting in a severe deficiency in the response speed and adaptability of existing systems to unexpected operating conditions. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: an intelligent operation and maintenance system for prefabricated substations based on big data, comprising:

[0005] The data access unit is used to align the electrical operation data, environmental monitoring data and historical operation and maintenance records collected in the prefabricated substation with a unified time reference, and to construct a time-series data structure containing node operating status and associated constraint relationships based on the equipment connection topology.

[0006] The state evolution modeling unit receives time-series data structures, extracts features at multiple time scales, and constructs a set of state evolution parameters that reflect the equipment's operating trends and potential anomaly evolution processes.

[0007] The control variable reconstruction unit receives a set of state evolution parameters and constructs a mapping relationship corresponding to real-time control variables. This mapping relationship consists of a set of component mapping functions that project each state evolution parameter into a control variable correction value based on its physical properties. These functions project each state evolution parameter into a corresponding control variable correction value based on its physical properties, and then superimpose these correction values ​​according to the device connection topology to form a correction parameter set. The control variable reconstruction unit then synthesizes the correction parameter set with real-time measurements in the same dimension to generate an enhanced control input vector.

[0008] The real-time control execution unit generates reactive power regulation and voltage control commands based on the enhanced control input vector and outputs them to the execution equipment of the prefabricated substation.

[0009] The human-machine feedback unit is used to introduce human correction information during the generation of the state evolution parameter set. It adjusts the parameters in the mapping relationship to change the generation result of the correction parameter set and writes the generation result back to the control variable reconstruction unit to update the composition of the enhanced control input vector.

[0010] Among them, the time-series data structure refers to the data sequence after unified time alignment and propagation based on device connection topology constraints; the state evolution parameter set refers to the combination of device operating parameters composed of impedance evolution parameters, capacity evolution parameters, and load evolution parameters.

[0011] As a further technical solution, the specific construction method of the time-series data structure is as follows: First, the analog quantities such as three-phase voltage, three-phase current, active power, and reactive power in the electrical operation data are converted into floating-point numerical sequences. The specific conversion rules are as follows: For the original analog quantity at each sampling time, the unit is first normalized, where voltage is based on the nominal voltage, current on the rated current, and power on the rated capacity; then multiplied by the range scaling factor and mapped to the [0,1] interval to form a floating-point numerical sequence; wherein, the range scaling factor is determined as follows: take the upper limit of the maximum allowable measurement range of the analog quantity in actual operation and perform the same unit normalization, the range scaling factor = 1 ÷ the maximum value of the normalization; according to the industry standard source and engineering application scenario, the upper limit of the maximum measurement range of voltage is determined to be 1.2 times the nominal voltage, corresponding to a range scaling factor of 1 / 1.2; the upper limit of the maximum measurement range of current is determined to be 1.5 times the rated current, corresponding to a range scaling factor of 1 / 1.5; the upper limit of the maximum measurement range of power is determined to be 1.0 times the rated capacity, corresponding to a range scaling factor of 1.0.

[0012] The temperature and humidity values ​​in the environmental monitoring data are segmented and quantized according to the product of temperature and humidity to obtain an integer level sequence. The segmented quantization is specifically performed by calculating the product P of temperature and relative humidity. temp_hum Five level ranges are set: P temp_hum ∈[0,500) corresponds to level 0, [500,1000) corresponds to level 1, [1000,2000) corresponds to level 2, [2000,3500) corresponds to level 3, and [3500,+∞) corresponds to level 4. The level values ​​are integers. The time difference Δt between the fault occurrence time in the historical maintenance record and the current time is exponentially decayed and then binarized into an event flag. Specifically, the decay function is defined. Where τ is the time constant, with a value of 720 hours, and a binarization threshold θ = 0.1 is set. When f(Δt) > θ, the event flag is set to 1, otherwise it is set to 0. Then, using the association constraint relationship of each edge in the equipment connection topology as the verification benchmark, the electrical operation data sequence is used as the main variable, the environmental level sequence is used as the weighting coefficient, and the event flag is used as the trigger mask, and they are all input into the constraint propagation function constructed based on the topological adjacency matrix. The constraint propagation function adopts the residual propagation rule of the topological edge weight coupling of the transformer substation, where the topological edge weight is defined as the reciprocal of the normalized impedance magnitude of the line connecting two nodes. The larger the edge weight, the tighter the electrical coupling. The method for calculating the residual vector is as follows: for each node i, the measured voltage value V in its electrical operation data is taken as the value of the residual vector. i Compared with the expected voltage value calculated by the power flow equation The difference is used as the residual component. The residual propagation rule is as follows: First, calculate the residual vector r of the electrical operation data of each node, and then use the weighting coefficient w. env (i.e., environmental level sequence value) multiplied by the inverse proportionality coefficient of topological edge weights To correct the magnitude of the residual, where α is an adjustable coefficient, w edge The edge weights between the current node and its neighboring nodes are defined. A trigger mask (event flag, 0 or 1) is used to select whether to accumulate the corrected residuals to the data sequence of neighboring nodes in the direction of decreasing topological edge weights. The decreasing direction is defined as from the node with the larger edge weight to the node with the smaller edge weight. All node data processed by the constraint propagation function are concatenated in topological order, which is defined as the breadth-first traversal order starting from the root node (the incoming node of the high-voltage side of the transformer). An integrity check header is attached to the beginning of the concatenation result. This integrity check header consists of the cumulative XOR value of all historical operation and maintenance record event flags. The cumulative XOR value is calculated by arranging all event flags in ascending order of timestamp and performing XOR operations sequentially, with an initial value of 0. All data blocks are linked one after another in ascending time order. The header of the next block contains the hash reduction value of the integrity check header of the previous block. The hash reduction value is calculated using the SHA-256 hash function to obtain a 256-bit hash value from the integrity check header of the previous block, and the first 32 bits are taken as the reduction value.

[0013] As a further technical solution, the state evolution parameter set includes impedance evolution parameters, capacity evolution parameters, and load evolution parameters. The specific calculation method for the impedance evolution parameters is as follows: An adaptive sliding window is applied to the voltage-to-current ratio sequence (V / I sequence e) of the same node at different timestamps in the time-series data structure. The window length of the adaptive sliding window is dynamically adjusted according to the local rate of change of the sequence. The initial window length is set to 10 sampling points. The standard deviation σ of the data within the window is calculated. If σ is greater than a set threshold of 0.05, the window length is shortened to 5 points; otherwise, the window length is extended to 20 points. A first-order linear regression model is used for fitting to obtain the linear trend slope of the ratio over time. The temperature value from environmental monitoring data is then introduced as a compensation factor. The compensation method is as follows: [Final impedance evolution parameters...] Where β is the temperature compensation coefficient, taken as 0.003 / ℃, and T is the current temperature. ref The reference temperature is 25℃; the calculation method for capacity evolution parameters is as follows: for the electrical operation data sequence related to reactive power output capability in the time-series data structure, specifically: the maximum reactive power increment ΔQ provided by the node. max Perform an exponentially weighted moving average (EWMA) with a smoothing coefficient λ of 0.2. The recursive formula is as follows: When a fault occurs in the historical maintenance log, the C at that moment will be recorded. evo Reset to initial value C evo (0) = 0.5 (per unit); The calculation method for the load evolution parameter is as follows: Perform multi-scale wavelet decomposition on the instantaneous sampled value sequence of node currents in the time series data structure, where the wavelet basis function is db4, the decomposition level is 4, the low-frequency approximation coefficients of the 4th level are extracted and reconstructed to obtain the low-frequency trend component, which is used as the load evolution parameter L. evo The three parameters are constructed using voltage, current, and reactive power data within the same time window in the time-series data structure as the starting point for calculation. The time window width is 15 minutes, and the sliding step size is 5 minutes. A three-parameter topology linkage correction mechanism is adopted: when the change in impedance evolution parameter ΔZ... evo When the absolute difference between the current value and the previous window value exceeds a set threshold of 0.05 (per unit), the change is multiplied by the damping coefficient ξ of the capacity evolution parameter. c =0.3 multiplied by the topological edge weight w edge and the inertia coefficient ξ of the load evolution parameter. l =0.4 multiplied by the topology split ratio s ratio The product is defined as the ratio of the node current to the sum of the currents of all adjacent nodes. The resulting product is then added to the current values ​​of the capacity evolution parameter and the load evolution parameter according to the topological node adjacency relationship. The addition method is used for the addition.

[0014] As a further technical solution, the following calculation method is specifically adopted when generating the correction values ​​for each control variable: the impedance evolution parameter Z is... evo The value of the voltage regulation correction ΔV is obtained by multiplying the node voltage V (measured in real time) by the node current I (measured in real time), then multiplying the result by the average impedance evolution parameters of all adjacent nodes in the device connection topology, and finally filtering with a first-order hysteresis filter. reg The time constant T of the first-order lag filter f Take 0.5s, sampling period T s =1s, first-order lag filter coefficient α=T s / (T s +T f ) = 1 / (1+0.5) = 0.6667; the recursive formula is: ; The capacity evolution parameter C evo Multiply by the reactive power deficit ΔQ in the real-time measurement. def It is defined as the target reactive power minus the current reactive power, or 0 if negative, and then divided by the time interval T between the two most recent reactive power compensation actions in the historical maintenance record. interval With rated response time T rate The ratio of T to T rate Take 2 seconds, and after limiting, use it as the reactive power compensation correction amount ΔQ. comp The limiting range is set to [-0.2, 0.2] (per unit); the load evolution parameter L is set to... evo Within the sliding time window, the average value over a 10-minute window is subtracted from the load evolution parameter value at the start of the window. This is then multiplied by the difference between the maximum and minimum fluctuations of the load evolution parameter within the window (i.e., the difference between the peak and trough values). After dead-zone clamping, this is used as the load prediction correction ΔL. pred The dead-zone clamping rule is: if the absolute value is less than 0.01, output 0; otherwise, keep the original value.

[0015] As a further technical solution, the component mapping function group enhances the correction accuracy through iterative feedback, specifically by performing the following steps: first, the impedance evolution parameter Z... evo Multiplying the node voltage V in the real-time measurement by the node current I yields the first intermediate quantity M1. This first intermediate quantity is then multiplied by the median of the impedance evolution parameters of all adjacent nodes in the device connection topology. When the number of adjacent nodes is even, the arithmetic mean of the two median values ​​is taken. After first-order hysteresis filtering, the result is used as the initial value ΔV for the voltage regulation correction. reg0 Simultaneously, the capacity evolution parameter C evo The reactive power deficit ΔQ in real-time measurements def Multiply and then divide by the ratio of the time interval between the two most recent reactive power compensation actions in the historical maintenance records to the rated response time, R=Tinterval / T rate Then multiply this ratio by the sign function of the initial value of the voltage regulation correction and output sign(ΔV). reg0 When ΔV reg0 When ΔV > 0, sign = 1, ΔV reg0 When <0, sign=-1, ΔV reg0 When = 0, sign = 0; thus, the initial value ΔQ of the reactive power compensation correction is obtained. comp0 Subsequently, the load evolution parameter L evo The initial value ΔL for the load prediction correction is obtained by subtracting the load evolution parameter value at the beginning of the window from the current sampled value within the sliding time window, and then multiplying it by the difference between the maximum and minimum fluctuation amplitudes of the load evolution parameters within the window after dead-zone clamping. pred0 Then, the initial value of the load prediction correction is multiplied by the first feedback coefficient k according to the device connection topology. fb1 This is then superimposed onto the current value of the impedance evolution parameter as a new input to the linear constraint function, multiplied by the second feedback coefficient k. fb2 This value is then superimposed onto the current value of the capacity evolution parameter as a new input to the proportional constraint function, where the expression for the linear constraint function is: The expression for the proportional constraint function is: ; using the updated Z evo_new and C evo_new Recalculate the updated values ​​of voltage regulation correction, reactive power compensation correction, and load prediction correction in sequence; perform iterative calculations until the absolute difference between the voltage regulation correction values ​​obtained from two adjacent iterations is less than the preset convergence threshold ε=0.001 (per unit value); then, define the three correction values ​​obtained from the last iteration as the weighted sum of the normalized reciprocal of the sum of the edge weights of the node and all its adjacent nodes in the device connection topology, and form a set of correction parameters.

[0016] As a further technical solution, the generation process of the enhanced control input vector includes the following sub-steps: adjusting the voltage regulation correction amount ΔV in the correction parameter set... reg Reactive power compensation correction amount ΔQ comp and load prediction correction ΔL pred The initial vector P of each node is obtained by summing the adjacency weights of the nodes in the device connection topology after normalization. init The normalized summation method is to multiply each of the three correction values ​​of a node by its adjacent weights, sum them, and then divide by the sum of all adjacent weights of that node; using the node voltage V in the real-time measurement as the main reference variable, the initial correction vector P is... initAs the object to be adjusted, the input constraint propagation function (the same as the constraint propagation function defined in the data access unit, but the main input variable is changed to the initial vector, while the weighting coefficients and trigger mask remain unchanged) is used. The integer level sequence corresponding to the environmental monitoring data is used as the weighting coefficient, and the event flags corresponding to the historical maintenance records are used as the trigger mask. The voltage residual vector of each node is calculated using the method P. ini The difference between the component in t and the expected value of the node voltage is propagated along the topological edge weights to obtain the voltage correction sequence V. corr After normalizing each component in the sequence according to its node adjacency weight, multiply it element-wise with the corresponding component of the initial modified vector to obtain the coupling correction C. coup Then, the initial correction vector is accumulated back according to the topological adjacency relationship to update its value. The accumulation rule is that for each node, the coupling correction of all its adjacent nodes is multiplied by the edge weight between the node and its adjacent nodes, the sum is multiplied by a decay factor of 0.5, and then added to the initial correction vector component of the node. This process is repeated until the L2 norm change of two adjacent voltage correction sequences is less than the preset convergence threshold ε. vec =0.0005, thus obtaining the final modified vector P. final The ratio of the absolute difference between two consecutive voltage adjustment corrections to the preset convergence threshold ε is used as the adaptive fusion coefficient. The adaptive fusion coefficient is multiplied by the sign function of the final correction vector (components are sign-based), and then multiplied element-wise by the final correction vector to obtain the scaling correction vector P. scale Finally, the real-time measurements of voltage, current, and power of each node are unified in terms of dimensions before synthesis. The current and power are converted to per-unit values ​​according to their rated values ​​and added element by element to generate an enhanced control input vector.

[0017] The human-machine feedback unit receives the expected offset direction and offset amount for the load prediction correction from the manual correction information. The expected offset direction is a discrete value: +1 indicates an increase in expectation; or -1 indicates a decrease in expectation. The offset amount Δ... exp The value is a unitless real number, ranging from -0.2 to 0.2. The direction factor is obtained by normalizing this offset using the reciprocal of the median of the load evolution parameters of adjacent nodes in the device connection topology. When the median is zero, take 1; then extract the actual value ΔL of the load prediction correction in the current correction parameter set. pred_actual Compared with the expected value ΔL pred_desired deviation Multiply the deviation by the sign of the current rate of change of the impedance evolution parameter (the rate of change is defined as the difference between the current window value and the previous window value divided by the sampling interval, and its sign is taken) and the sign of the current rate of change of the capacity evolution parameter, respectively, to obtain the first feedback adjustment amount. Second feedback adjustment amount The first feedback adjustment amount is added to the first feedback coefficient, i.e. The summed values ​​are limited to the range [0.05, 0.3]; the second feedback adjustment is then added to the second feedback coefficient, i.e. The feedback coefficients are limited to the range [0.1, 0.4] and the iteration calculation is repeated until the absolute difference between the voltage regulation corrections obtained from two adjacent iterations is less than half of the preset convergence threshold ε. The final set of correction parameters is output as the generation result, and the updated first and second feedback coefficients are written back to the mapping relationship storage area in the control variable reconstruction unit. This storage area saves the historical values ​​of the feedback coefficients of each node in the form of key-value pairs.

[0018] As a further technical solution, the human-machine feedback unit will update the first feedback coefficient k. fb1_new With the second feedback coefficient k fb2_new Multiply by the current rate of change dZ of the impedance evolution parameter respectively evo / dt and the current rate of change of the capacity evolution parameter dC evo / dt, to obtain the first modulation factor Second modulation factor The first modulation factor is used to perform weighted correction on the integer level sequence in the constraint propagation function. The correction method is to multiply the original integer level sequence value by (1+m1), then round it down and limit it to the [0,4] interval. The second modulation factor is used to perform amplitude modulation on the trigger mask of the event flag. The modulation method is to multiply the original Boolean event sequence by (1+m2), and then map it to the [0,1] interval through the Sigmoid function as a continuous trigger mask. The control variable reconstruction unit re-executes the iterative adjustment from the initial correction vector to the final correction vector with the corrected integer level sequence and the modulated event flag, and uses the product of the first modulation factor and the second modulation factor as the exponent of the exponential decay base of the adaptive fusion coefficient. The exponential decay base is the natural constant e. The specific update formula is as follows: , where γ old For adaptive fusion coefficients, exp is an exponential function with the natural constant e as the base; using the updated γ new The scaling correction vector is recalculated and then added element-wise to the real-time measurements to update the composition of the enhanced control input vector.

[0019] This invention provides an intelligent operation and maintenance system for prefabricated substations based on big data, which has the following beneficial effects:

[0020] 1. This invention constructs a time-series data structure by unifying time alignment and topological constraints of multi-source heterogeneous data, thereby improving the integrity of transformer substation data fusion and the degree of matching of topological constraints.

[0021] 2. This invention solves the problems of difficulty in capturing state trends and lack of physical interpretability of control mapping by using three-parameter state evolution modeling and physical attribute component mapping iteration, thereby improving the accuracy of operation and maintenance prediction and the adaptive capability of control.

[0022] 3. This invention solves the problem of the difficulty in integrating human experience into automated control by manually correcting the driving parameters iteratively, writing them back, and deeply embedding them into the entire control process, thereby improving the system's adaptability to operating conditions and the flexibility of operation and maintenance decisions. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] refer to Figure 1 This embodiment takes an outdoor unattended 10kV / 0.4kV prefabricated substation with a rated capacity of 1000kVA as the application object, and fully implements the entire operation process of the intelligent operation and maintenance system for prefabricated substations based on big data described in this invention. The system includes five core units: data access unit, state evolution modeling unit, control variable reconstruction unit, real-time control execution unit, and human-machine feedback unit. Those skilled in the art can directly reproduce the system functions completely according to the steps, data, and parameter standards of this embodiment, and realize the full-process intelligent operation and maintenance of prefabricated substations, including multi-source data fusion, state trend modeling, adaptive control, and human experience embedding. It effectively solves the core problems in the prior art, such as lack of topological constraints in multi-source data, difficulty in describing abnormal propagation paths, inability to continuously embed human experience into automated control, and insufficient adaptability to unexpected operating conditions.

[0026] After the system starts up, it first enters the workflow of the data access unit. This unit first performs a unified time reference alignment operation on all data collected in the prefabricated substation, sets a unified sampling period of 1 second, and ensures that electrical operation data, environmental monitoring data, and historical operation and maintenance records are collected and transmitted synchronously on the same time axis. The electrical operation data includes three-phase voltage, three-phase current, active power, and reactive power analog quantities on the high-voltage and low-voltage sides of the prefabricated substation. The environmental monitoring data is the real-time temperature and relative humidity inside the enclosed cabinet of the prefabricated substation. The historical operation and maintenance records retrieve the fault occurrence records, reactive power compensation device operation records, and on-load tap changer operation records of the past 30 days.

[0027] After time alignment, the data access unit constructs a time-series data structure based on the transformer substation equipment connection topology, including node operating status and associated constraints. This topology uses the high-voltage side incoming node of the transformer substation as the root node, and sequentially connects the transformer node, low-voltage bus node, and each outgoing circuit node. The topology edge weight is defined as the reciprocal of the normalized impedance magnitude of the line connecting two nodes; a larger edge weight indicates a tighter electrical coupling relationship between the two nodes. During the construction of the time-series data structure, the three types of data are first converted into sequence data in their corresponding formats. Electrical operating data is converted into floating-point numerical sequences. During the conversion, unit normalization and range scaling rules are strictly followed, and the voltage is based on the 10kV nominal voltage. Based on the standard, the upper limit of the maximum measurement range is set at 1.2 times the nominal voltage, i.e., 12kV, with a range scaling factor of 1 / 1.2. For current, based on the rated current of 57.7A on the high-voltage side and 1443A on the low-voltage side, the upper limit of the maximum measurement range is set at 1.5 times the rated current, with a range scaling factor of 1 / 1.5. For power, based on the rated capacity of 1000kVA, the upper limit of the maximum measurement range is set at 1.0 times the rated capacity, with a range scaling factor of 1.0. All analog quantities are converted and mapped to the [0,1] interval to form a standard floating-point sequence. Environmental monitoring data is converted to an integer-level sequence. The internal temperature of the transformer is 28℃ and the relative humidity is 60% in real time. The temperature and humidity product P is calculated. temp_hum=28×60=1680. Following the preset rules that [0,500) corresponds to level 0, [500,1000) to level 1, [1000,2000) to level 2, [2000,3500) to level 3, and [3500,+∞) to level 4, it is quantized to an integer level 2. If the ambient temperature rises to 35℃ and the humidity to 80%, the product 2800 corresponds to level 3. Quantization is strictly performed according to the preset interval throughout the process. Historical maintenance records are converted into Boolean event sequences, and the time difference Δt=4 between the most recent fault occurrence and the current time is extracted. For an 8-hour period, with a time constant τ = 720 hours and a binarization threshold θ = 0.1, the attenuation function f(Δt) ≈ 0.9355 is calculated. If this value is greater than the threshold, the event flag is set to 1. If the time difference Δt = 1650 hours, the calculated attenuation function result is approximately 0.098, which is less than the threshold, and the event flag is set to 0. After completing the data format conversion, the association constraint relationship of each edge in the equipment connection topology is used as the verification benchmark. The electrical operation data sequence is used as the main variable, the environmental level sequence is used as the weighting coefficient, and the event flag is used as the trigger mask. These are all input into the approximate value constructed based on the topological adjacency matrix. In the constraint propagation function, the residual propagation rule of the topology weight coupling of the transformer substation is adopted. For example, the adjustable coefficient α is set to 0.5. First, the residual vector between the measured voltage value of each node and the expected voltage of the power flow equation is calculated. Then, the residual magnitude is corrected by multiplying the weighting coefficient by the inverse proportional coefficient of the topology weight. The trigger mask is used to select whether to accumulate the corrected residual to the data sequence of the adjacent node along the weight decreasing direction. The weight decreasing direction is defined as from the node with larger weight to the node with smaller weight. All node data after being processed by the constraint propagation function are strung together in breadth-first traversal order starting from the root node. The concatenation process involves appending an integrity check header to the beginning of the concatenation result. This header consists of the cumulative XOR values ​​of all historical operation and maintenance record event flags. The initial XOR value is 0. After performing XOR operations on event flags 1, 0, and 1 in sequence, the final check header is 1. All data blocks are linked sequentially in ascending chronological order. The header of each subsequent block contains the hash reduction value of the integrity check header of the previous block. The hash reduction value is calculated using the SHA-256 hash function to obtain a 256-bit hash value from the integrity check header of the previous block. The first 32 bits are then used as the reduction value. This completes the construction of a time-series data structure with topological constraints and time correlation.

[0028] The process then proceeds to the state evolution modeling unit. This unit receives the time-series data structure output from the data access unit, extracts multi-timescale features, and constructs a state evolution parameter set consisting of impedance evolution parameters, capacity evolution parameters, and load evolution parameters. The impedance evolution parameters are obtained by adaptively sliding window fitting the voltage-current ratio sequence at different timestamps of the same node in the time-series data structure. The initial window length of the adaptive sliding window is set to 10 sampling points. The standard deviation of the data within the window is calculated to be σ=0.03, which is less than the preset threshold of 0.05. The window length is automatically extended to 20 sampling points. A first-order linear regression model is used to obtain the linear trend slope of the ratio over time. Then, a temperature compensation factor is introduced, with the temperature compensation coefficient β set to 0.003 / ℃ and the reference temperature T. ref =25℃, current temperature 28℃, the final impedance evolution parameter is the slope multiplied by (1+0.003×(28-25)); the capacity evolution parameter is calculated using an exponentially weighted moving average based on the maximum reactive power increment sequence related to reactive power output capacity in the time series data structure, with a smoothing coefficient λ of 0.2, and the parameter values ​​are updated in real time according to the recursive formula. When a fault occurs in the historical operation and maintenance record, the capacity evolution parameter is immediately reset to the initial per-unit value of 0.5; the load evolution parameter is decomposed into 4 levels using the db4 wavelet basis for the instantaneous sampling value sequence of node current in the time series data structure, the low-frequency approximation coefficients of the 4th level are extracted and reconstructed, and the resulting low-frequency trend component is the load evolution parameter; all three parameters are calculated based on voltage, current, and reactive power data within a time window with a width of 15 minutes and a sliding step of 5 minutes, and a three-parameter topology linkage correction mechanism is enabled. When the impedance evolution parameter changes by ΔZ evo When the value of 0.06 (per unit) exceeds the preset threshold of 0.05, the change is multiplied by the product of the capacity evolution parameter damping coefficient 0.3 and the topology edge weight, and the product of the load evolution parameter inertia coefficient 0.4 and the topology current split ratio. The topology current split ratio is defined as the proportion of the node current to the sum of the currents of all adjacent nodes. The current is superimposed on the current values ​​of the capacity evolution parameter and the load evolution parameter according to the adjacency relationship of the topology nodes, and finally a set of state evolution parameters that can accurately reflect the equipment operation trend and potential abnormal evolution process is formed.

[0029] Next, the workflow of the control variable reconstruction unit begins. This unit receives the set of state evolution parameters and projects each parameter into control variable correction quantities according to its physical properties using a set of component mapping functions. These correction parameters are then iteratively fed back and superimposed on the topology to form a set of correction parameters. Finally, this set is synthesized with the real-time measurements in the same dimension to generate an enhanced control input vector. First, the initial values ​​of the three types of control variable correction quantities are calculated. The impedance evolution parameters are multiplied by the node voltage in the real-time measurements, and then divided by the node current to obtain the first intermediate quantity. This intermediate quantity is then multiplied by the median of the impedance evolution parameters of all adjacent nodes of that node, and filtered with a first-order lag of 0.5 seconds to obtain the voltage. Adjust the initial value of the correction; multiply the capacity evolution parameter by the reactive power deficit in the real-time measurement, divide by the ratio of the time interval between the two most recent reactive power compensation actions to the rated response time of 2 seconds, multiply by the sign function of the initial value of the voltage regulation correction to obtain the initial value of the reactive power compensation correction, with the limiting range set to [-0.2, 0.2] per unit; subtract the window start value from the current sampled value of the load evolution parameter within the 10-minute sliding window, multiply by the difference between the maximum and minimum fluctuation amplitudes within the window, and after dead-zone clamping with an absolute value less than 0.01, obtain the initial value of the load prediction correction; then, adjust the initial value of the load prediction correction. The values ​​are multiplied by the initial first feedback coefficient of 0.1 and the initial second feedback coefficient of 0.2, respectively, and the impedance evolution parameters and capacity evolution parameters are updated and iteratively calculated again. The calculation continues until the absolute difference between the voltage regulation correction values ​​obtained from two adjacent iterations is less than the preset convergence threshold of 0.001 per unit. The three correction values ​​obtained from the last iteration are weighted and summed according to the node adjacency weights to form a correction parameter set. In the enhanced control input vector generation stage, the correction parameter set is normalized and summed according to the node adjacency weights to obtain the initial correction vector. The real-time node voltage is used as the main reference variable. The initial correction vector is input into the constraint propagation function and the environment is shaped. The system uses the grade sequence as a weighting coefficient and historical maintenance event flags as trigger masks to calculate the voltage residual vector and propagate it along the topological edge weights. This process is repeated iteratively until the L2 norm change of two adjacent voltage correction sequences is less than the preset convergence threshold of 0.0005, resulting in the final correction vector. The ratio of the absolute difference between two adjacent voltage adjustment corrections to the preset convergence threshold is used as the adaptive fusion coefficient. This coefficient is multiplied by the sign function of the final correction vector and then multiplied element-wise with the final correction vector to obtain the scaling correction vector. Finally, this vector is added element-wise with the unified dimension real-time measurement to complete the generation of the enhanced control input vector, providing accurate input basis for subsequent control execution.

[0030] The system then enters the workflow of the real-time control execution unit. This unit analyzes the control parameters based on the enhanced control input vector output by the control variable reconstruction unit, and generates reactive power regulation and voltage control commands in combination with the characteristics of the transformer substation execution equipment. These commands are then output to the reactive power compensation device, on-load tap changer, and other execution equipment. In this embodiment, when the enhanced control input vector detects that the low-voltage side voltage is 0.02 per unit low and the reactive power deficit is 0.15 per unit, the system immediately generates a voltage regulation command to drive the on-load tap changer to increase one level. At the same time, it generates a reactive power compensation command to connect a set of capacitors. The command response time strictly matches the rated response time of 2 seconds. After the execution equipment acts quickly, the low-voltage side voltage of the transformer substation returns to the rated value, and the reactive power reaches a balanced state. This effectively avoids voltage fluctuations and reactive power losses, ensures the continuous and stable operation of the transformer substation, and achieves precise adaptive control of reactive power and voltage.

[0031] Finally, the workflow of the human-machine feedback unit is initiated. This unit embeds manual correction information during system operation, achieving a deep integration of operational experience and automated control. Operational personnel input the desired offset direction for the load prediction correction amount via the human-machine interface, which is an increase (discrete value + 1), and the offset amount Δ. exp =0.1 per unit value. The system normalizes the offset by the reciprocal of the median of the load evolution parameters of adjacent nodes to obtain the direction factor. The median of the load evolution parameters of adjacent nodes is 0.2, and the direction factor is calculated as 0.1 / 0.2=0.5. The deviation e=0.05 between the actual value and the expected value of the current load prediction correction is extracted. The deviation is multiplied by the sign of the current rate of change of the impedance evolution parameter and the capacity evolution parameter, respectively, to obtain the first feedback adjustment and the second feedback adjustment. The first feedback adjustment is added to the initial first feedback coefficient of 0.1, limiting... After the value is updated to 0.15 in the interval [0.05, 0.3], the second feedback adjustment is added to the initial second feedback coefficient of 0.2. After being restricted to the interval [0.1, 0.4], the value is updated to 0.25. The new feedback coefficient is used to recalculate iteratively until the absolute difference between the voltage regulation correction values ​​obtained from two adjacent iterations is less than half of the convergence threshold, i.e., 0.0005. The final correction parameter set is then output, and the updated feedback coefficients are written back to the mapping relationship storage area of ​​the control variable reconstruction unit in the form of key-value pairs to save the historical values ​​of the feedback coefficients of each node.

[0032] Furthermore, the human-machine feedback unit multiplies the updated first and second feedback coefficients by the current rate of change of the impedance evolution parameter and the capacity evolution parameter, respectively, to obtain the first and second modulation factors. The first modulation factor is used to weight and correct the integer level sequence in the constraint propagation function. The original environmental level 2 is corrected and rounded to 2, with the amplitude limited to the [0,4] interval. The second modulation factor is used to modulate the amplitude of the event flag trigger mask and map it to a continuous trigger mask through the Sigmoid function. Based on the corrected environmental level sequence and the modulated trigger mask, the control variable reconstruction unit re-executes the iterative adjustment from the initial correction vector to the final correction vector. The product of the first and second modulation factors is used as the exponent of the base of the adaptive fusion coefficient exponential decay. After updating the adaptive fusion coefficient, the scaling correction vector is recalculated and synthesized with the real-time measurement to update the composition of the enhanced control input vector. This allows the field experience of the operation and maintenance personnel to be continuously embedded in the dynamic generation process of control parameters, realizing the system's rapid response and adaptive optimization to unexpected operating conditions.

[0033] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart operation and maintenance system for prefabricated substations based on big data, characterized in that, include: The data access unit is used to align the electrical operation data, environmental monitoring data and historical operation and maintenance records collected in the prefabricated substation with a unified time reference, and to construct a time-series data structure containing node operating status and associated constraint relationships based on the equipment connection topology. The state evolution modeling unit receives time-series data structures, extracts features at multiple time scales, and constructs a set of state evolution parameters that reflect the equipment's operating trends and potential anomaly evolution processes. The control variable reconstruction unit receives a set of state evolution parameters and constructs a mapping relationship corresponding to the real-time control variables. This mapping relationship is a set of component mapping functions that project each state evolution parameter into a control variable correction value based on its physical properties. These functions project each state evolution parameter into a corresponding control variable correction value based on its physical properties, and then superimpose these correction values ​​according to the equipment connection topology to form a correction parameter set. The control variable reconstruction unit synthesizes the correction parameter set with the real-time measured quantities in the same dimension to generate an enhanced control input vector. The equipment connection topology refers to the electrical connection relationship between the high-voltage side incoming line, transformer, low-voltage busbar, and outgoing circuit of the prefabricated substation, using a breadth-first topology structure with the high-voltage side incoming line as the root node. The real-time control execution unit generates reactive power regulation and voltage control commands based on the enhanced control input vector and outputs them to the execution equipment of the prefabricated substation. The human-machine feedback unit is used to introduce manual correction information during the generation of the state evolution parameter set. It adjusts the parameters in the mapping relationship to change the generation result of the correction parameter set, and writes the generation result back to the control variable reconstruction unit to update the composition of the enhanced control input vector.

2. The intelligent operation and maintenance system for prefabricated substations based on big data as described in claim 1, characterized in that: The time-series data structure includes: converting electrical operation data, environmental monitoring data, and historical maintenance records into floating-point numerical sequences, integer level sequences, and Boolean event sequences, respectively. Environmental monitoring data is quantized into integer levels by segmenting the numerical values ​​based on the product of temperature and humidity. Historical maintenance records are binarized into event flags based on the exponential decay of the time difference between the fault occurrence time and the current time. Then, using the association constraint relationship of each edge in the device connection topology as the verification benchmark, the electrical operation data sequence is used as the main variable, the environmental level sequence as the weighting coefficient, and the event flags as the trigger mask, all of which are input into a constraint propagation function constructed based on the topological adjacency matrix. In this process, the constraint propagation function adopts the residual propagation rule of topology weight coupling of the transformer substation: First, the residual vector of electrical operation data of each node is calculated. Then, the magnitude of the residual is corrected by multiplying the weighting coefficient by the inverse proportional coefficient of the topology weight. The trigger mask is used to select whether to accumulate the corrected residual to the data sequence of the adjacent node in the direction of decreasing topology weight. All node data after being processed by the constraint propagation function are concatenated in topological order, and an integrity check header consisting of the cumulative XOR value of all historical operation and maintenance record event flags is appended to the beginning of the concatenation result. All data blocks are linked one after another in ascending time order, and the header of the next block contains the hash reduction value of the integrity check header of the previous block.

3. The intelligent operation and maintenance system for prefabricated substations based on big data as described in claim 1, characterized in that: The state evolution parameter set includes impedance evolution parameters, capacity evolution parameters, and load evolution parameters. Impedance evolution parameters are obtained by adaptively sliding window fitting the voltage-to-current ratio sequence of the same node at different time stamps in the time-series data structure and incorporating temperature values ​​from environmental monitoring data as a compensation factor. Capacity evolution parameters are obtained by performing an exponentially weighted moving average on the electrical operation data sequence related to reactive power output capacity in the time-series data structure and resetting the weighting coefficients using the fault occurrence times in historical maintenance records. Load evolution parameters are obtained by... The instantaneous sampled value sequence of the current is decomposed by multi-scale wavelet decomposition and low-frequency trend components are extracted. The three parameters are constructed using voltage, current and reactive power data within the same time window in the time-series data structure as the calculation starting point, and a three-parameter topology linkage correction mechanism is adopted: when the change of the impedance evolution parameter exceeds the set threshold, the change is multiplied by the damping coefficient of the capacity evolution parameter multiplied by the topology edge weight, and the inertia coefficient of the load evolution parameter multiplied by the topology shunting ratio, and the resulting product is superimposed on the current values ​​of the capacity evolution parameter and the load evolution parameter according to the adjacency relationship of the topology nodes.

4. The intelligent operation and maintenance system for prefabricated substations based on big data according to claim 3, characterized in that: The impedance evolution parameter is multiplied by the node voltage in the real-time measurement and then divided by the node current in the real-time measurement. The resulting ratio is then multiplied by the average impedance evolution parameter of all adjacent nodes in the equipment connection topology. After first-order hysteresis filtering, this value is used as the voltage regulation correction. The capacity evolution parameter is multiplied by the reactive power deficit in the real-time measurement and then divided by the ratio of the time interval between the two most recent reactive power compensation actions in the historical operation and maintenance records to the rated response time. After limiting, this value is used as the reactive power compensation correction. The average load evolution parameter within the sliding time window is subtracted from the load evolution parameter value at the beginning of the window. Then, this value is multiplied by the difference between the maximum and minimum fluctuation amplitudes of the load evolution parameter within the window. After dead-zone clamping, this value is used as the load prediction correction.

5. The intelligent operation and maintenance system for prefabricated substations based on big data as described in claim 1, characterized in that: The component mapping function group includes: first, multiplying the impedance evolution parameter by the node voltage in the real-time measurement and dividing by the node current to obtain a first intermediate value; then multiplying this first intermediate value by the median of the impedance evolution parameters of all adjacent nodes in the device connection topology, and then filtering by first-order hysteresis; the result is used as the initial value of the voltage regulation correction; simultaneously, multiplying the capacity evolution parameter by the reactive power deficit in the real-time measurement and dividing by the ratio of the time interval between the two most recent reactive power compensation actions in the historical maintenance record to the rated response time; then multiplying this ratio by the sign function of the initial value of the voltage regulation correction to obtain the initial value of the reactive power compensation correction; subsequently, subtracting the load evolution parameter value at the beginning of the window from the current sampled value of the load evolution parameter within the sliding time window, and then multiplying by the value within the window. The difference between the maximum and minimum fluctuation amplitudes of the load evolution parameters, after dead-zone clamping, is used as the initial value of the load prediction correction. This initial value is then multiplied by the first feedback coefficient according to the device connection topology and superimposed on the current value of the impedance evolution parameter as a new input to the linear constraint function. It is also multiplied by the second feedback coefficient and superimposed on the current value of the capacity evolution parameter as a new input to the proportional constraint function. The updated values ​​of the voltage regulation correction, reactive power compensation correction, and load prediction correction are recalculated sequentially. Iterative calculations are performed until the absolute difference between two adjacent voltage regulation corrections is less than a preset convergence threshold. Finally, the three corrections obtained in the last iteration are weighted and summed according to the node adjacency weights in the device connection topology to form a set of correction parameters.

6. The intelligent operation and maintenance system for prefabricated substations based on big data according to claim 5, characterized in that: The voltage regulation correction, reactive power compensation correction, and load prediction correction in the correction parameter group are normalized and summed according to the node adjacency weight in the equipment connection topology to obtain the initial correction vector of each node. Using the node voltage in real-time measurements as the main reference variable, the initial correction vector is taken as the object to be adjusted. The constraint propagation function is input, and the integer level sequence corresponding to the environmental monitoring data is used as the weighting coefficient and the event flag corresponding to the historical operation and maintenance record is used as the trigger mask. The voltage residual vector of each node is calculated and propagated along the topological edge weight to obtain the voltage correction sequence. After normalizing each component in the sequence according to the node adjacency weight, it is multiplied element by element with the corresponding component of the initial correction vector to obtain the coupling correction. Then, it is accumulated back into the initial correction vector according to the topological adjacency relationship to update its value. Repeat the iteration until the L2 norm change of two adjacent voltage correction sequences is less than the preset convergence threshold to obtain the final correction vector; use the ratio of the absolute difference between two adjacent voltage adjustment corrections to the preset convergence threshold as the adaptive fusion coefficient, multiply the adaptive fusion coefficient by the sign function of the final correction vector, and then multiply it element by element with the final correction vector to obtain the scaling correction vector. Finally, add it element by element with the real-time measurement to generate the enhanced control input vector.

7. The intelligent operation and maintenance system for prefabricated substations based on big data as described in claim 6, characterized in that: The human-machine feedback unit receives the expected offset direction and offset amount for the load prediction correction amount from the manual correction information. It normalizes the offset amount by the reciprocal of the median of the load evolution parameters of adjacent nodes in the device connection topology to obtain the direction factor. Then, it extracts the deviation between the actual value of the load prediction correction amount and the manual expected value in the current correction parameter set. It multiplies the deviation by the sign of the current rate of change of the impedance evolution parameter and the sign of the current rate of change of the capacity evolution parameter to obtain the first feedback adjustment amount and the second feedback adjustment amount. The first feedback adjustment amount is superimposed on the first feedback coefficient, and the second feedback adjustment amount is superimposed on the second feedback coefficient. The feedback coefficient is updated and the iterative calculation is re-executed until the absolute difference of the voltage regulation correction amount obtained by two adjacent iterations is less than half of the preset convergence threshold. The final correction parameter set is output as the generation result, and the updated first feedback coefficient and second feedback coefficient are written back to the mapping relationship storage area in the control variable reconstruction unit.

8. The intelligent operation and maintenance system for prefabricated substations based on big data according to claim 7, characterized in that: The human-machine feedback unit multiplies the updated first feedback coefficient and second feedback coefficient by the current rate of change of the impedance evolution parameter and the capacity evolution parameter, respectively, to obtain the first modulation factor and the second modulation factor. The first modulation factor is used to perform weighted correction on the integer level sequence in the constraint propagation function, and the second modulation factor is used to perform amplitude modulation on the trigger mask of the event flag. The control variable reconstruction unit re-executes the iterative adjustment from the initial correction vector to the final correction vector with the corrected integer level sequence and the modulated event flag, and uses the product of the first modulation factor and the second modulation factor as the exponent of the exponential decay base of the adaptive fusion coefficient. After recalculating the scaling correction vector, it adds it element-wise with the real-time measurement quantity to update the composition of the enhanced control input vector.

Images