Electric power communication private network management system based on 5G slicing technology

Abstract

The invention provides an electric power communication private network management system based on a 5G slicing technology, relates to the technical field of electric power communication, breaks an information barrier between a communication network and a power grid, synchronously acquires power grid side electrical quantity original data and network side channel state data through an edge computing gateway, and performs timestamp alignment and multi-dimensional feature fusion. And deducing the fused space-time state tensor by using a deep neural network, and predicting the burst probability of generating high-quality services (such as relay protection tripping) in a future short time slot in advance. Based on the prediction result, the system dynamically generates a resource reservation mask on a physical layer, and preemptive allocation of time-frequency resource blocks is completed before a key instruction is generated, so that zero-latency transmission of high-reliability services is realized, and the deterministic low-delay requirements of power production control services are ensured.

Description

Technical Field

[0001] This application relates to the field of power communication technology, and more specifically, to a power communication private network management system based on 5G slicing technology. Background Technology

[0002] With the deepening construction of the energy internet and new power systems, power communication private networks are facing the dual challenges of heterogeneous service models and massive terminal access. To simultaneously support high-bandwidth services such as transmission and transformation inspection video and ultra-reliable, low-latency services such as distribution automation and relay protection on a unified physical infrastructure, a power communication private network management system based on 5G network slicing technology has emerged. This technology aims to achieve flexible resource isolation and on-demand allocation through network soft slicing, and is a key means to solve the last-mile access problem of the power grid.

[0003] However, in actual power grid production control scenarios, existing slice resource management technologies still have significant limitations. Current 5G base station schedulers primarily employ a reactive scheduling mechanism based on buffer states, meaning the system only begins calculating and allocating resources after high-priority data packets arrive at the base station buffer pool. This delayed processing method leads to critical control commands often being forced to queue when sudden power grid faults occur (such as short-circuit tripping) because the wireless spectrum is already filled with background video streams. The resulting queuing delay easily exceeds the millisecond-level hard real-time constraints of relay protection services, thus triggering the risk of cascading power grid faults. The root cause lies in the information barrier between the communication network and the power grid physical system; the communication scheduler cannot perceive the precursors of transient fluctuations in physical quantities such as power grid voltage and current, thus failing to achieve predictive resource reservation. Furthermore, in environments with strong electromagnetic interference, such as substations, the operation of high-voltage equipment (such as the opening and closing of disconnecting switches) generates unevenly distributed pulse noise in the frequency domain. Existing technologies typically employ continuous frequency band optimization strategies, which can easily cause reserved resource blocks to fall precisely within high-noise interference bands, resulting in the failure of high-priority service transmission. Meanwhile, traditional resource preemption methods often blindly punch holes, ignoring the system bit protection when background services (such as eMBB) use LDPC encoding. Frequent destruction of key codewords leads to a large number of retransmissions of background services, which seriously reduces the overall spectrum efficiency of the system.

[0004] Therefore, an optimized power communication private network management system based on 5G slicing technology is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a power communication private network management system based on 5G slicing technology, which includes: The data acquisition module is used to synchronously acquire raw grid-side data containing instantaneous values ​​of three-phase voltage and current, as well as raw network-side data containing slice buffer queue length and channel quality indication, in real time through the edge computing gateway. The data alignment module is used to align the timestamps of the raw data from the power grid side and the raw data from the network side to obtain a joint raw dataset. The feature fusion module is used to normalize the joint original dataset and extract high-frequency transient component features, and then concatenate the extracted feature vector with the network state vector in multiple dimensions to obtain a fusion state tensor that represents the comprehensive panoramic state of the power grid and network at the current moment. The prediction and extrapolation module is used to predict and extrapolate the burst probability of high-optimal services on the fusion state tensor to obtain the burst probability value. The resource decision module is used to determine the resource block allocation matrix to guide the medium access control layer in resource scheduling based on the burst probability value and the channel quality data in the joint original dataset. The scheduling execution module is used to control the base station scheduler to perform slice-level resource allocation based on the resource block allocation matrix, and to prioritize the transmission of high-priority service data packets on the reserved time and frequency resource blocks.

[0006] Compared with existing technologies, the power communication private network management system based on 5G slicing technology provided in this application synchronously collects instantaneous values ​​of electrical quantities on the power grid side and slice status data on the network side through an edge computing gateway, and constructs a fusion tensor representing the comprehensive panoramic status of the power grid and network using timestamp alignment technology. Based on this, a deep learning model is used to perform time-series extrapolation of high-frequency transient components, predicting the probability of bursts of high-priority services within short time slots in the future. Based on this prediction result, the system dynamically generates resource reservation masks at the physical layer, and forcibly locks time-frequency resource blocks before fault symptoms appear but before service packets arrive. This mechanism achieves a leap from reactive scheduling to predictive scheduling, effectively solving the problem of uncontrollable transmission delays caused by resource preemption delays in high-priority services under mixed service scenarios, and ensuring the deterministic transmission of power control commands. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1 This is a system block diagram of a power communication private network management system based on 5G slicing technology according to an embodiment of this application; Figure 2 This is a schematic diagram of data flow in a power communication private network management system based on 5G slicing technology according to an embodiment of this application; Figure 3 This is a block diagram of a feature fusion module in a power communication private network management system based on 5G slicing technology according to an embodiment of this application; Figure 4 This is a block diagram of the prediction and extrapolation module in the power communication private network management system based on 5G slicing technology according to an embodiment of this application; Figure 5 This is a block diagram of the resource decision module in a power communication private network management system based on 5G slicing technology according to an embodiment of this application; Figure 6 This is a block diagram of a matrix determination unit in a power communication private network management system based on 5G slicing technology according to an embodiment of this application. Detailed Implementation

[0009] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0010] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0011] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0012] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0013] Existing power communication private network management based on 5G slicing technology mainly relies on buffer states for reactive resource scheduling. This leads to critical control commands often being queued during sudden grid failures because wireless resources are already occupied by high-traffic background services, failing to meet millisecond-level hard real-time constraints. Furthermore, because the communication network cannot detect precursory voltage and current fluctuations at the grid's physical layer, existing technologies struggle to accurately reserve resources within the extremely short window before a fault occurs. Therefore, this application proposes a power communication private network management system based on 5G slicing technology. This system synchronously collects instantaneous values ​​of electrical quantities from the power grid side and slice status data from the network side through an edge computing gateway. Utilizing timestamp alignment and multi-dimensional feature extraction techniques, it constructs a fused state tensor containing high-frequency transient characteristics of electrical components and network congestion trends. Subsequently, a deep learning model is used to perform time-series extrapolation on this tensor, accurately predicting the probability of bursts of high-priority services within short time slots in the future, and dynamically generating a resource block allocation matrix at the physical layer accordingly. Finally, the system controls the base station scheduler to lock in preferred time-frequency resource blocks before service packets arrive, achieving a leap from delayed response to proactive prediction in scheduling, effectively ensuring deterministic transmission and low-latency requirements for critical power services in mixed service scenarios.

[0014] Figure 1 This is a system block diagram of a power communication private network management system based on 5G slicing technology according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in a power communication private network management system based on 5G slicing technology according to an embodiment of this application. Figure 1 and Figure 2 As shown, the power communication private network management system 100 based on 5G slicing technology according to an embodiment of this application includes: a data acquisition module 110, used to synchronously acquire grid-side raw data containing instantaneous values ​​of three-phase voltage and current and network-side raw data containing slice buffer queue length and channel quality indication in real time through an edge computing gateway; a data alignment module 120, used to perform timestamp alignment on the grid-side raw data and the network-side raw data to obtain a joint raw dataset; and a feature fusion module 130, used to perform normalization processing and high-frequency transient component feature extraction on the joint raw dataset, and to fuse the extracted feature vector with... The network state vector is multidimensionally concatenated to obtain a fused state tensor representing the current overall state of the power grid and network; the prediction and inference module 140 is used to predict and infer the burst probability of high-priority services on the fused state tensor to obtain the burst probability value; the resource decision module 150 is used to determine the resource block allocation matrix to guide the medium access control layer to perform resource scheduling based on the burst probability value and the channel quality data in the joint original dataset; the scheduling execution module 160 is used to control the base station scheduler to perform slice-level resource allocation based on the resource block allocation matrix, and prioritize the transmission of high-priority service data packets on the reserved time-frequency resource blocks.

[0015] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the data acquisition module 110 is used to synchronously acquire, in real time, raw grid-side data containing instantaneous values ​​of three-phase voltage and current, and raw network-side data containing slice buffer queue length and channel quality indication through an edge computing gateway. It should be noted that, in traditional power communication network management, an information barrier exists between the communication network and the power grid physical system. The base station scheduler cannot detect subtle fluctuations in grid voltage or current, leading to queuing delays in communication resources during critical moments of fault occurrence due to the inability to reserve them in advance. Based on this, the technical solution of this application first synchronously acquires, in real time, raw grid-side data containing instantaneous values ​​of three-phase voltage and current, and raw network-side data containing slice buffer queue length and channel quality indication through an edge computing gateway. Through this processing, the data isolation between the power grid and the communication network can be effectively broken down, providing a high-quality data foundation with strict time alignment for subsequent multimodal feature fusion and predictive resource scheduling, ensuring the accuracy of cross-domain data analysis.

[0016] More specifically, in a concrete example of this application, asynchronous buffering and parsing of heterogeneous data streams are first performed. The industrial bus interface of the edge computing gateway receives real-time bitstreams from distribution automation terminals or phasor measurement units, and performs application-layer data unit parsing to extract the instantaneous values ​​of three-phase voltage and three-phase current. Simultaneously, a key performance indicator data stream containing slice buffer queue length and channel quality indication is received through the base station's northbound interface, and its abstract syntax markup is parsed. Next, for cases where the sampling rates of the two data streams are inconsistent, independent circular buffers are established for temporary storage, and data frames with checksum errors are removed. Subsequently, time base unification and time stamp extraction operations are performed. An absolute timestamp based on satellite time synchronization is extracted from the raw data from the power grid side, and the system frame number and subframe number based on a precise clock protocol are extracted from the raw data from the network side. The relative time on the network side is uniformly converted to a Coordinated Universal Time (UTC) nanosecond-level timestamp using a local reference clock, thereby eliminating time discrepancies between heterogeneous systems. Finally, a sliding window-based resampling and timing alignment operation is performed. The sampling time axis is constructed with the network-side scheduling cycle as the master clock reference. A small tolerance window is set before and after the master clock point. For the original grid-side data with a high sampling rate, the electrical quantity values ​​at the master clock time are calculated using zero-order hold or linear interpolation methods to ensure that the electrical characteristics and network state correspond to the same physical time. Finally, a joint original dataset that has been time-synchronized and dimension-stitched is generated.

[0017] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the data alignment module 120 is used to perform timestamp alignment between the original data from the power grid side and the original data from the network side to obtain a joint original dataset. It should be noted that since the data from the power grid side and the data from the network side originate from heterogeneous systems with drastically different physical mechanisms, they naturally exhibit asynchrony in sampling frequency, clock source, and reporting period. Direct feature splicing would lead to state misalignment and fail to accurately reflect the temporal causal relationship between voltage fluctuations and network congestion. Therefore, the technical solution of this application further performs timestamp alignment between the original data from the power grid side and the original data from the network side to obtain a joint original dataset. Through this processing, the time deviation between multi-source heterogeneous data can be eliminated, ensuring that electrical physical quantities and network performance indicators strictly correspond under the same time reference, providing a reliable data foundation for the subsequent construction of a high-precision fused state tensor.

[0018] More specifically, in a specific example of this application, the data alignment module 120 includes: an asynchronous buffering unit, used to asynchronously buffer and unpack the original data from the power grid side and the original data from the network side based on a dual-channel ring buffering mechanism to obtain heterogeneous buffered data pairs; a time base unification unit, used to unify the time base and extract the time stamps from the heterogeneous buffered data pairs to obtain a standardized time-series data stream with a unified time dimension; and a resampling alignment unit, used to resample and align the standardized time-series data stream based on a sliding window to obtain a joint original dataset.

[0019] More specifically, the process begins with asynchronous buffering and protocol unpacking based on a dual-channel ring buffer mechanism. Independent ring buffers are established for temporary storage of the two data streams. The application layer data units of the power grid data and the encoding formats of the network data are parsed separately, and data frames with verification anomalies are removed to obtain heterogeneous buffered data pairs. Next, time base unification and time stamp extraction are performed. Absolute timestamps based on satellite time synchronization are extracted from the power grid data, and system frame numbers and subframe numbers based on a precise clock protocol are extracted from the network data. The relative time on the network side is uniformly converted to Coordinated Universal Time (UTC) nanosecond-level timestamps using a local reference clock, thereby generating a standardized time-series data stream with a unified time dimension. Finally, a sliding window-based resampling and time alignment operation is performed. A sampling time axis is constructed using the network side's scheduling cycle as the master clock reference. Small tolerance windows are set before and after the master clock point. For power grid data with a high sampling rate, an interpolation algorithm is used to calculate the master clock value to obtain the joint original dataset. In this process, the calculation of the kth joint data sample after alignment satisfies a specific mathematical relationship, namely, the value of the master clock alignment time is estimated based on the discrete original sampling points of the power grid through the interpolation function, and the sampling point that is closest to the master clock time in the actual recorded data of the network side is selected, and the time difference between the two must be less than the maximum allowable time jitter tolerance threshold.

[0020] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the feature fusion module 130 is used to normalize the joint original dataset and extract high-frequency transient component features. The extracted feature vector is then multidimensionally concatenated with the network state vector to obtain a fused state tensor representing the current overall state of the power grid and network. It should be noted that, given the fundamental differences in physical dimensions and numerical magnitude between the electrical physical quantities on the power grid side and the communication state parameters on the network side in the joint original dataset, direct fusion would lead to difficulty in model training convergence. Furthermore, early signs of power grid faults are often hidden in the weak high-frequency transient components of voltage or current waveforms, and simple time-domain numerical values ​​are insufficient to carry enough fault prediction information. Based on this, the technical solution of this application further normalizes the joint original dataset and extracts high-frequency transient component features. Max-min normalization and Z-Score standardization are used to eliminate dimensional differences between heterogeneous data. The short-time Fourier transform algorithm is used to map the electrical data to the time-frequency domain to capture high-frequency transient features. Simultaneously, the growth rate trend of the network buffer is calculated. The extracted feature vectors and network state vectors are then multidimensionally concatenated and time-series sliding window stacked along the feature dimension to obtain a fused state tensor representing the current comprehensive panoramic state of the power grid and network. Through the above processing, a high-dimensional feature space containing both the physical micro-fluctuations of the power grid and the congestion trend of the communication network can be effectively constructed. This allows for in-depth mining of the nonlinear spatiotemporal correlations between cross-domain data, providing highly robust and discriminative input features for subsequent high-priority service burst probability prediction.

[0021] Figure 3 This is a block diagram of a feature fusion module in a power communication private network management system based on 5G slicing technology, according to an embodiment of this application. Figure 3 As shown, the feature fusion module 130 includes: a normalization processing unit 131, used to normalize the electrical data and network data in the joint original dataset to obtain normalized data pairs; a feature extraction unit 132, used to extract high-frequency transient and trend features from the normalized data pairs to obtain a heterogeneous feature vector set; and a sliding window stacking unit 133, used to slide window stack the multimodal features in the heterogeneous feature vector set in the time step dimension to obtain a fused state tensor.

[0022] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the normalization processing unit 131 is used to normalize the electrical data and network data in the joint original dataset to obtain normalized data pairs. It should be noted that, due to the fundamental differences in physical dimensions and numerical magnitudes between the grid-side electrical physical quantities and network-side communication state parameters in the joint original dataset, directly using them for subsequent model training would cause the neural network to have difficulty converging during gradient descent, and features with larger values ​​could easily mask features with smaller values ​​but containing crucial information. Therefore, the technical solution of this application further normalizes the electrical data and network data in the joint original dataset to obtain normalized data pairs. Through this processing, the dimensional influence between heterogeneous data can be effectively eliminated, mapping multimodal data to a unified numerical range, thereby improving the accuracy of subsequent feature extraction and the stability of model training.

[0023] More specifically, in a concrete example of this application, the joint original dataset is first split into an electrical subset and a network subset. For the voltage and current data in the electrical subset, considering the definite physical operating extremes of power equipment, a minimum-max normalization method is used to map them to a specific interval to preserve the relative amplitude characteristics of the waveforms. For the buffer length and channel quality indicators in the network subset, considering the statistical randomness of network traffic, a Z-score normalization method is used to eliminate the influence of outliers on the distribution. In this process, for the electrical data... and network data Calculate the normalized values ​​respectively. and The calculation process satisfies the following mathematical relationship: in, This represents the normalized electrical quantity value. This represents the input raw electrical sample value. and These represent the physical extreme constants of electrical equipment, such as the lower and upper percentage limits of the rated voltage. This represents the standardized values ​​of key network performance indicators. This represents the mean of the network metric within the sliding time window. This represents the standard deviation of the network metric within the sliding time window. This represents a small constant used to prevent the denominator from being zero.

[0024] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the feature extraction unit 132 is used to perform dual-track extraction of high-frequency transient and trend features on normalized data pairs to obtain a heterogeneous feature vector set. It should be noted that early signs of power grid faults are often hidden in the weak harmonic oscillations of voltage or current waveforms, and the network state values ​​at a single point in time cannot objectively reflect the dynamic evolution trend of traffic congestion. If only the original time-domain data is used for analysis, it is difficult to capture key abrupt changes. Based on this, the technical solution of this application further performs dual-track extraction of high-frequency transient and trend features on normalized data pairs to obtain a heterogeneous feature vector set. Through the above processing, the physical feature dimensions of the two types of data can be effectively decoupled, capturing high-frequency abrupt changes on the power grid side and the load accumulation rate on the network side respectively, providing high-resolution input variables for subsequent time-series prediction.

[0025] More specifically, in a particular example of this application, a short-time Fourier transform is performed on the normalized voltage and current sequences. A sliding window function is used to map the one-dimensional time-domain signal to the two-dimensional time-frequency domain, thereby extracting high-frequency transient component features that cannot be represented by a single time point value. Simultaneously, a first-order difference calculation is performed on the network buffer length sequence to obtain the buffer growth rate feature that characterizes the rate of traffic growth, rather than focusing solely on the current static queue length. In this process, electrical features... With network trend characteristics The calculations follow a specific mathematical model, where electrical characteristics are derived from normalized electrical signals. With window function The product of the complex exponential terms is accumulated to obtain the value, which represents the time... and frequency The amplitude of time-frequency characteristics at a given time; network trend characteristics are calculated by measuring the amplitude of time-frequency characteristics at the current time. Compared to the previous moment Normalized network metrics The difference between the sampling time interval and the sampling time interval The ratio is worth getting.

[0026] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the sliding window stacking unit 133 is used to perform sliding window stacking of multimodal features in the heterogeneous feature vector set along the time step dimension to obtain a fused state tensor. It should be noted that, since the evolution of power grid faults and the backlog of communication network traffic both have strong temporal dependencies, isolated features at a single moment cannot fully depict the dynamic causal process from the appearance of electrical symptoms to the occurrence of communication congestion, and the independence of heterogeneous data in physical space hinders in-depth mining of cross-domain nonlinear correlations. Based on this, the technical solution of this application further performs sliding window stacking of multimodal features in the heterogeneous feature vector set along the time step dimension to obtain a fused state tensor. Through the above processing, a high-order data structure containing historical time-series information and multidimensional modal features can be effectively constructed, providing a panoramic input with spatiotemporal correlation characteristics for subsequent deep learning models, thereby improving the accuracy and robustness of predicting the probability of high-priority service bursts.

[0027] More specifically, in a concrete example of this application, a dimension flattening operation is first performed on the feature vectors, reshaping the two-dimensional electrical time-frequency feature matrix generated in the previous steps into a one-dimensional high-dimensional vector to adapt to the subsequent splicing dimension requirements. Next, a time-series sliding window stacking operation is performed, introducing a historical observation window of a preset length, and simultaneously extracting the electrical feature vectors and network feature vectors of the current moment and several past moments according to the time step sequence. Subsequently, a tensor splicing operation of multimodal features is performed, concatenating the processed electrical feature sequence with the network state feature sequence in the feature dimension to construct a three-dimensional tensor with a shape adapted to the input interface of the deep neural network. During this process, for each time step… Fusion state tensor Its construction follows a specific mathematical model, namely .in, This represents the final generated fusion state tensor. This represents the vector concatenation operator. Indicates the length of the time step in the historical observation window. Indicates from time At that time Electrical characteristic sequence, This represents the network state feature sequence corresponding to the same period.

[0028] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the prediction and deduction module 140 is used to predict and deduce the burst probability of high-priority services from the fused state tensor to obtain the burst probability value. It should be noted that, given that traditional communication scheduling mechanisms mainly rely on the posterior feedback of buffer queue states, they exhibit significant response lag when dealing with sudden power grid faults. Furthermore, there is a complex nonlinear spatiotemporal coupling relationship between the micro-fluctuations of power grid electrical quantities and the macro-congestion of network traffic. Simple linear threshold determination is insufficient to accurately capture fleeting fault symptoms before the arrival of service packets. Based on this, the technical solution of this application further predicts and deduces the burst probability of high-priority services from the fused state tensor to obtain the burst probability value. It utilizes a deep recurrent neural network model loaded with pre-trained weights to perform forward propagation calculations on the input spatiotemporal tensor, deeply extracts implicit temporal dependency features through a gating mechanism, and combines an attention mechanism to weighted focus on key fault symptom time steps. Finally, through a nonlinear mapping layer, it outputs the conditional probability of high-priority services occurring within a preset time slot in the future. Through the above processing, the scheduling logic can be effectively transformed from passive response to proactive prediction. The transmission demand can be perceived at the physical symptom stage before the generation of millisecond-level power grid fault commands, thereby providing a forward-looking decision-making basis for subsequent resource reservation and ensuring deterministic low-latency transmission of power control commands in mixed business congestion scenarios.

[0029] Figure 4 This is a block diagram of the prediction and extrapolation module in a power communication private network management system based on 5G slicing technology, according to an embodiment of this application. Figure 4 As shown, the prediction and inference module 140 includes: a deep extraction unit 141, used to perform time-series feature deep extraction on the fused state tensor based on a recurrent neural network to obtain a hidden layer feature sequence; an attention focusing unit 142, used to perform attention-weighted focusing on the hidden layer feature sequence at key time steps to obtain a context semantic vector focused on fault symptoms; and a probability mapping unit 143, used to perform nonlinear probability mapping and confidence output on the context semantic vector to obtain a burst probability value.

[0030] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the deep extraction unit 141 is used to perform time-series feature deep extraction on the fused state tensor based on recurrent neural networks to obtain a hidden layer feature sequence. It should be noted that, since the evolution of power grid faults and the backlog state of communication network buffers both exhibit strong time-varying characteristics and long-range dependencies, static feature analysis at a single moment is insufficient to capture the dynamic evolution trend of fault precursors and cannot effectively establish a causal relationship between historical electrical fluctuations and current network congestion. Based on this, the technical solution of this application further performs time-series feature deep extraction on the fused state tensor based on recurrent neural networks to obtain a hidden layer feature sequence. Through the above processing, the gating mechanism can be effectively used to filter and memorize key historical state information, and the nonlinear coupling characteristics of multimodal data in the time dimension can be mined, thereby providing a hidden layer expression containing complete contextual semantics for subsequent fault probability inference.

[0031] More specifically, in a concrete example of this application, pre-trained prediction model weights are first loaded into a pre-defined deep neural network architecture, which employs a bidirectional long short-term memory network to simultaneously capture forward and backward temporal dependencies. Subsequently, the fused state tensor is injected as the input sequence into the network for forward propagation computation, recursively processing the feature vector at each time step along the time axis, and utilizing the forgetting gate, input gate, and output gate mechanisms within the network to regulate the forgetting and accumulation of information. During this process, for each time step... Input Generate hidden state vector Its calculation process satisfies a specific mathematical model: in, Indicates at time step The hidden state vector output by the unit at this moment This represents the activation function used to generate the gating signal. It represents the Hadamah accumulation. The first term in the input fusion state tensor represents the... Feature slices at each time step , , This represents the weight matrix and bias term of the forget gate. , , The weight matrix and bias terms represent the cell state update.

[0032] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the attention focusing unit 142 is used to perform attention-weighted focusing on the hidden layer feature sequence at key time steps to obtain a context semantic vector focused on fault symptoms. It should be noted that, in the long-term sequence of power grid fault evolution, key symptoms indicating fault occurrence often only exist within extremely short transient time windows, such as the instant of voltage drop or current change, while the data at most other time steps are steady-state background noise. If equal weights are assigned to the features at all time steps, the key fault precursor information will be diluted by the stable data, thereby reducing the sensitivity of the prediction model. Based on this, the technical solution of this application further performs attention-weighted focusing on the hidden layer feature sequence at key time steps to obtain a context semantic vector focused on fault symptoms. Through the above processing, the focusing mechanism of human vision can be effectively simulated, automatically identifying and strengthening the features of abrupt changes that contribute significantly to the prediction results, while suppressing noise interference from irrelevant time steps. This compresses the variable-length time-series information into a high-density feature vector representing fault semantics, ensuring that the model can accurately capture fleeting fault symptoms.

[0033] More specifically, in a concrete example of this application, the attention score calculation operation is first performed. A learnable query vector and weight parameters are introduced to perform a non-linear mapping on the hidden state at each time step in the sequence, calculating an energy score representing the importance of that moment. Next, a weight normalization operation is performed, using an exponential function to convert the energy scores of all time steps into attention weight coefficients in probability distribution form, ensuring that the sum of the weights is one. Subsequently, a weighted summation and semantic compression operation is performed, linearly combining the hidden state vectors of all time steps according to the calculated weight coefficients to generate a fixed-dimensional context semantic vector. In this process, the computational logic follows a specific mathematical definition: first, through the weight matrix... Bias terms and projection vector For time Hidden state The energy fraction is obtained by transformation. ,Right now Secondly, using the exponential function right Attention weights are obtained by normalization. ,Right now Finally, calculate the hidden states at all time points. With corresponding weights The sum of the products is used to obtain the context vector. ,Right now .

[0034] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the probability mapping unit 143 is used to perform nonlinear probability mapping and confidence output on the context semantic vector to obtain the burst probability value. It should be noted that, since the context semantic vector, after being focused by the attention mechanism, is essentially still an implicit expression in a high-dimensional abstract feature space, it cannot be directly used as the quantitative basis for the medium access control layer scheduler to execute binary resource reservation decisions. Furthermore, simple linear mapping is difficult to adapt to the nonlinear distribution characteristics of the probability of power grid fault occurrence, easily leading to blurred decision boundaries. Based on this, the technical solution of this application further performs nonlinear probability mapping and confidence output on the context semantic vector to obtain the burst probability value. Through the above processing, high-dimensional fault symptom features can be compressed and mapped to a closed interval between zero and one, generating a conditional probability scalar with clear physical meaning. This provides accurate confidence input for subsequent hysteresis comparison and resource block calculation, ensuring that resource preemption operations are triggered only when the fault risk is extremely high, avoiding the waste of spectrum resources due to false alarms.

[0035] More specifically, in a concrete example of this application, a fully connected mapping operation is first performed, inputting the context semantic vector into the multilayer perceptron classification head. A learnable weight matrix and bias term are used to project the high-dimensional semantic vector onto a one-dimensional real space to integrate the contributions of different feature dimensions to fault occurrence. Subsequently, nonlinear activation and probability normalization operations are performed, using the Sigmoid activation function to compress the projected real value, strictly limiting its output to the range of zero to one. This output value represents the confidence level of high-priority services generated on the power grid side within a preset time slot in the future. During this process, the burst probability value... The calculations follow a specific nonlinear mapping model: in, This represents the burst probability value of the final output. This represents the contextual semantic vector of the input. and These represent the weight matrix and bias term of the fully connected output layer, respectively. Since is a natural constant, this nonlinear transformation achieves a smooth transition from the feature space to the decision probability space.

[0036] Specifically, considering the non-stationary dynamic changes in power system load fluctuations and communication network traffic characteristics over time, this application constructs a model maintenance mechanism combining offline training based on historical backtracking and incremental updates based on online data. In the offline supervised training phase, the fused state tensor after normalization and feature extraction within a historical time period is selected as the input sample. Whether high-priority service messages were actually detected within a preset time window after the corresponding moment is used as the true supervision label, constructing a training dataset containing positive and negative samples. The binary cross-entropy loss between the predicted probability and the true label is calculated using the backpropagation algorithm, and the weight matrix and bias terms in the deep recurrent neural network are iteratively updated using the adaptive moment estimation optimization algorithm. In the online incremental fine-tuning phase, the measured input and service arrival results of the most recent period are cached in real time. When the accumulated prediction error exceeds a preset threshold or reaches a predetermined time period, the model parameters are updated using small-step gradient descent updates based on newly collected measured data. During this process, the target loss function... The calculations follow a specific mathematical model: in, This indicates the number of samples in the training batch. Indicates the first The true label of each sample The representation model is for the first The predicted probability value of each sample output. This represents the set of weight parameters in the network that need to be updated.

[0037] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the resource decision module 150 is used to determine a resource block allocation matrix to guide the medium access control layer in resource scheduling based on burst probability values ​​and channel quality data in the joint original dataset. It should be noted that, given that simple fault prediction probability values ​​cannot be directly recognized as executable physical resource allocation instructions by the medium access control layer scheduler at the base station level, and that the time-varying characteristics of wireless channels determine that the number of time-frequency resources required to transmit the same size power control message differs under different signal quality conditions, the technical solution of this application further determines a resource block allocation matrix to guide the medium access control layer in resource scheduling based on burst probability values ​​and channel quality data in the joint original dataset. It converts continuously changing probability values ​​into discrete preemption trigger flags through a hysteresis comparison mechanism, and searches for the corresponding spectral efficiency based on real-time channel quality indicators to calculate the number of physical resource blocks required to transmit the target payload, thereby generating a binary mask mapped onto a two-dimensional time-frequency grid. Through the above processing, a mapping bridge can be effectively established from high-level prediction logic to low-level physical control. While ensuring that the resources required for high-priority services are accurately locked, the reserved bandwidth can be dynamically adjusted using the channel adaptive mechanism to avoid spectrum waste caused by fixed reservation or transmission failure caused by insufficient reservation.

[0038] Figure 5 This is a block diagram of the resource decision module in a power communication private network management system based on 5G slicing technology according to an embodiment of this application. Figure 5 As shown, the resource decision module 150 includes: a threshold comparison unit 151, used to perform a hysteresis comparison between the burst probability value and preset high-water level thresholds and low-water level thresholds to obtain a preemption trigger flag indicating whether to perform resource reservation; a resource quantity calculation unit 152, used to find the corresponding spectral efficiency according to the channel quality indication in the joint original dataset when the preemption trigger flag is valid, and calculate the required frequency domain resource quantity for transmission in combination with the preset target payload size to obtain the required resource block count; and a matrix determination unit 153, used to determine the resource block allocation matrix based on the required resource block count.

[0039] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the threshold comparison unit 151 is used to perform a hysteresis comparison between the burst probability value and preset high-water level thresholds and low-water level thresholds to obtain a preemption trigger flag indicating whether to execute resource reservation. It should be noted that since the burst probability value output by the deep learning model is a continuously changing analog quantity, and may oscillate frequently near a single decision boundary due to small fluctuations in input characteristics, directly using single-threshold logic for binary judgment would cause frequent switching between the effective and ineffective states of the resource reservation instruction, thereby triggering a base station control signaling storm and disrupting scheduling stability. Based on this, the technical solution of this application further performs a hysteresis comparison between the burst probability value and preset high-water level thresholds and low-water level thresholds to obtain a preemption trigger flag indicating whether to execute resource reservation. Through the above processing, the hysteresis filtering characteristic triggered by Schmitt triggering can be effectively introduced, eliminating short-term random jitter of the predicted probability value, ensuring that resource preemption is triggered only when the fault risk certainty is high, and releasing resources only after the risk has significantly decreased, thereby ensuring the robustness and smoothness of power communication private network resource scheduling decisions.

[0040] More specifically, in a specific example of this application, a threshold parameter loading operation is first performed, reading the pre-configured high-water level threshold for determining the activation of preemption mode and the low-water level threshold for determining the deactivation of preemption mode. These two thresholds together define a state-maintaining hysteresis interval. Subsequently, a state transition determination operation for the current moment is performed, comparing the burst probability value obtained through real-time extrapolation with the aforementioned two thresholds to determine the flag state for the next moment. During this process, the time... Preemption trigger flag The generation strictly follows the following hysteresis comparison mathematical model: if and only if the burst probability value at the current time is Greater than the high water level threshold When the preemption trigger flag is set to logic 1, it indicates that resource reservation should be performed immediately; if and only if the current burst probability value is... Less than the low water level threshold When the preemption trigger flag is set to logic 0, it indicates that the resource lock should be released; if the burst probability value falls between the high-water mark threshold and the low-water mark threshold, the preemption trigger flag state from the previous moment should be forcibly maintained. Unchanged. It should be noted that at the initial system startup time (t=0), by default... Initialized to logic 0 (non-preemptive state) to avoid accidental triggering during cold starts. This logic enables denoising of high-priority service scheduling instructions, avoiding frequent oscillations in the physical resource block allocation matrix caused by minor fluctuations in model predictions near critical points.

[0041] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the resource quantity calculation unit 152 is used to find the corresponding spectral efficiency based on the channel quality indication in the joint original dataset when the preemption trigger flag is valid, and calculate the required frequency domain resources for transmission in combination with the preset target payload size to obtain the required resource block count. It should be noted that due to the high time-varying nature of wireless communication environments, the channel conditions of the same terminal at different times directly determine the amount of information that can be carried per unit bandwidth. Furthermore, although high-priority power control messages have small payloads, they have extremely high requirements for transmission reliability. If a fixed resource allocation method is used, transmission failure will occur when the channel quality is poor, while spectrum waste will occur when the channel quality is good. Based on this, the technical solution of this application further finds the corresponding spectral efficiency based on the channel quality indication in the joint original dataset when the preemption trigger flag is valid, and calculates the required frequency domain resources for transmission in combination with the preset target payload size to obtain the required resource block count. Through the above processing, adaptive scheduling of physical layer resources can be effectively realized, ensuring that the minimum frequency domain resources required by the current channel state are dynamically matched while meeting the high-reliability transmission target, thereby maximizing the system's spectrum utilization.

[0042] More specifically, in a concrete example of this application, a resource requirement calculation operation is first performed. The calculation process is initiated only when the preemption trigger flag is set to an active state, determining the target payload size based on the standard power control message format. Next, a spectrum efficiency mapping operation is performed. Based on the latest channel quality indication in the joint raw dataset, a pre-set modulation and coding strategy mapping table is consulted to obtain the number of bits that can be transmitted per symbol under the current channel conditions, i.e., the spectrum efficiency. Subsequently, a physical resource block quantization operation is performed. Taking into account the target payload size, spectrum efficiency, and control signaling overhead, the number of physical resource blocks required to carry the service is calculated. During this process, the required resource block count is performed. The calculations follow a specific mathematical model: in, This represents the final output count of resource blocks required. This indicates a preemption trigger flag, used to control the start and stop of the calculation process. This represents the floor function, ensuring that the allocated resource blocks are integers. This indicates the preset data packet size for high-priority power control commands. This represents the redundancy safety factor in resource allocation, used to offset the risk of channel fluctuations. This indicates the number of subcarriers contained in a physical resource block in the frequency domain. This represents the number of valid orthogonal frequency division multiplexing symbols within a scheduling time unit. This represents the spectral efficiency mapping function for channel quality indication.

[0043] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the matrix determination unit 153 is used to determine the resource block allocation matrix based on the required resource block count. It should be noted that, given that in scenarios with strong electromagnetic interference such as substations, frequency-selective impulse noise generated by disconnecting switch operations can cause specific frequency bands to remain in a high-interference state for extended periods, a simple continuous frequency band optimization strategy could easily result in the selected set of resource blocks falling entirely within the noise pollution zone. Furthermore, blindly seizing resources ignores the crucial role of system bit pair decoding when using low-density parity-check codes for background enhanced mobile broadband services, easily leading to direct verification failure of background service transmission blocks. Based on this, the technical solution of this application further determines the resource block allocation matrix based on the required resource block count, comprehensively utilizes the subcarrier signal-to-noise ratio data, the preset power impulse noise spectrum, and the background service coding mapping table in the joint original dataset, constructs a comprehensive utility function that includes the reliability of power grid perception and the cost of background service impairment, quantitatively evaluates each physical resource block within the full bandwidth, and performs a greedy optimization decision with frequency domain discrete constraints to obtain the optimal discrete frequency index set, ultimately mapping it to a comb-shaped punch mask on a time-frequency two-dimensional grid. Through the above processing, precise resource reservation can be effectively achieved at the physical layer to avoid strong electromagnetic interference frequency bands and minimize background service coding loss. While ensuring high-reliability transmission of power control commands, the overall spectrum efficiency and background service throughput of the system can be improved.

[0044] In one embodiment of this application, the matrix determination unit 153 includes: a region selection unit, used to perform region selection and mask marking in a time-frequency two-dimensional resource grid according to the required resource block count; and a state setting unit, used to set the elements in the selected region to a reserved state to obtain a resource block allocation matrix.

[0045] Furthermore, the initialization and region selection operations of the time-frequency two-dimensional resource grid are first performed to construct a zero-based matrix whose dimensions cover the total number of resource blocks and the number of symbols per time slot in the system. If the required resource block count is greater than zero, a starting position is determined on the frequency domain index axis according to the principle of frequency domain continuity or the principle of optimal channel quality, and several consecutive frequency indices starting from this starting position are selected as the preferred region. Subsequently, the state setting and matrix generation operations are performed, traversing each time-frequency element in the base matrix. The values ​​of elements located within the preferred region are modified to reserved state identifiers, while the values ​​of elements outside the preferred region are kept as non-reserved state identifiers, thereby generating the final resource block allocation matrix. In this process, this matrix serves as the scheduling instruction at the bottom layer of the physical layer, instructing the base station scheduler to perform puncturing or silent operations at specific time-frequency positions to reserve continuous physical channels for upcoming high-reliability services such as relay protection.

[0046] In another embodiment of this application, the inventors discovered that in environments such as substations, disconnector operation generates frequency-selective impulse noise. If the aforementioned continuous frequency band strategy is adopted, the selected resource block may fall entirely within the noise pollution band. Furthermore, to avoid damaging critical system bits of background eMBB services and causing decoding failure, this embodiment proposes a discrete resource optimization mechanism based on electromagnetic noise spectrum and minimum coding loss.

[0047] Figure 6 This is a block diagram of a matrix determination unit in a power communication private network management system based on 5G slicing technology according to a preferred embodiment of this application. Figure 6 As shown, the matrix determination unit 153 includes: a utility evaluation unit 1531, used to perform a comprehensive utility evaluation on each physical resource block within the full bandwidth based on the subcarrier signal-to-noise ratio data, a preset power impulse noise spectrum, and a background service coding mapping table in the joint original dataset to obtain a resource utility score vector; an optimal decision unit 1532, used to perform a greedy optimal decision with frequency domain discrete constraints on the resource utility score vector and the required number of physical resource blocks to obtain an optimal discrete frequency index set; and a mask mapping unit 1533, used to perform comb mask mapping processing on a time-frequency two-dimensional grid on the optimal discrete frequency index set to obtain a resource block allocation matrix.

[0048] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the utility evaluation unit 1531 is used to perform a comprehensive utility evaluation on each physical resource block within the full bandwidth based on the subcarrier signal-to-noise ratio data, a preset power impulse noise spectrum, and a background service coding mapping table in the joint original dataset to obtain a resource utility score vector. It should be noted that, given the complex electromagnetic environment of the power communication private network, frequency-selective impulse noise generated by substation disconnection switch operations or high-voltage corona discharge can cause a sharp deterioration in channel quality for specific frequency bands. Furthermore, blind resource grabbing often ignores the crucial role of system bit pair decoding when using low-density parity-check codes for background enhanced mobile broadband services, easily leading to transmission failures due to key bit loss in background services, thereby reducing overall spectral efficiency. Based on this, the technical solution of this application further performs a comprehensive utility evaluation on each physical resource block within the full bandwidth based on the subcarrier signal-to-noise ratio data, a preset power impulse noise spectrum, and a background service coding mapping table in the joint original dataset to obtain a resource utility score vector. Through the above processing, two core dimensions, namely power grid environment perception and service coding structure perception, can be effectively introduced to accurately identify the high-quality frequency points that are least affected by electromagnetic interference and have the least impact on the decoding of background services after being preempted. This ensures the reliable transmission of high-quality services while minimizing the bit error damage to background services.

[0049] More specifically, in a concrete example of this application, an operation is performed to construct a resource block utility and cost evaluation vector, which aims to quantify the availability and preemption cost of each physical resource block within the full bandwidth under the current power environment. The system receives a joint raw dataset containing real-time subcarrier signal-to-noise ratio, a power impulse noise map pre-obtained through field mapping or online learning, and an encoded mapping table of background services at the current moment. An index is then created for each physical resource block within the full bandwidth. We introduce a gain dimension based on the power impulse noise spectrum and a cost dimension based on low-density parity-check code coding mapping to calculate its comprehensive utility score. This score is directly proportional to the corrected channel capacity and inversely proportional to the risk of background service decoding loss caused by preemption of this position. The specific calculation logic follows the mathematical model below: in, For the first The comprehensive utility score of each resource block, the first term of the formula ( The corresponding power grid sensing reliability term represents the theoretical capacity of this resource block. For transmission power, For channel gain, Thermal noise power, The power of pulse noise interference unique to the power environment. This is the noise sensitivity coefficient, which ensures that frequency bands polluted by power grid noise are avoided. The second term in the formula ( The corresponding enhanced mobile broadband service impairment cost item represents the preemption cost. For indicator functions, if resource block The system bit that carries background services is 1, otherwise it is 0. Corresponding check bits, and These are the weights for the system bits and the parity bits, respectively. Since the probability of decoding failure due to the loss of system bits is much higher than that of the parity bits, weights are set accordingly. Much larger For example, take and , and This is the normalized weighting factor. In actual substation inspection scenarios, when high-priority relay protection commands need to preempt resources, the algorithm automatically avoids resource blocks that, although having a decent signal-to-noise ratio, are in the impulse noise frequency band. It also avoids resource blocks that carry critical system bits of the inspection video stream, prioritizing resource blocks that carry video stream verification bits and have a clean electromagnetic environment for transmission, thereby achieving the optimal trade-off for global performance.

[0050] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the preferred decision unit 1532 is used to perform a greedy optimal decision with frequency domain discrete constraints on the resource utility score vector and the required number of physical resource blocks to obtain a preferred discrete frequency index set. It should be noted that, given the complex electromagnetic environment of the power communication private network, in addition to narrowband noise in specific frequency bands, there are frequent unknown burst broadband interferences. If a traditional continuous resource allocation strategy is adopted, it is very easy for a selected set of resource blocks to be simultaneously covered by broadband interference, leading to complete transmission failure. Based on this, the technical solution of this application further performs a greedy optimal decision with frequency domain discrete constraints on the resource utility score vector and the required number of physical resource blocks to obtain a preferred discrete frequency index set. Through the above processing, the frequency domain diversity effect can be effectively utilized to resist broadband pulse interference. By forcibly constraining the frequency domain physical distance between the selected resource blocks, it is ensured that even if some frequency bands encounter burst interference, the remaining discretely distributed resource blocks can still maintain communication capabilities, thereby maximizing the comprehensive utility of the resource set while meeting the service transmission quantity requirements.

[0051] More specifically, in a concrete example of this application, a discrete optimization algorithm with a minimum frequency domain spacing constraint is executed. This algorithm aims to search for an optimal set of indexes across the entire bandwidth, such that the sum of the utility scores of all resource blocks in the set reaches the maximum value, and the frequency domain distance between any two selected resource blocks in the set is greater than a preset safety interval. This process is implemented by solving the constrained optimization problem using a heuristic greedy search, and the specific calculation formula is as follows: in, This is the optimal set of frequency indices, i.e., the preferred set of discrete frequency indices for the final output. The set of frequency indices represents the candidate subset in the search space. The number of resource blocks corresponding to the total bandwidth defines the physical scope of the search. This means maximizing the total utility of the selected set, that is, optimizing the set to have good channel quality, low noise interference, and minimal impact on background-enhanced mobile broadband services. This indicates that the maximum value should be returned. , The selected quantity must meet business needs, ensuring that the total reserved resources are sufficient to handle high-priority business data. To constrain frequency domain dispersion, the selected resource blocks are ensured to be discretely distributed in the frequency domain, forming a comb-like structure to avoid contiguous interference. Based on this, in scenarios where a high-voltage switch action in a substation generates a sudden pulse interference covering the width of several physical resource blocks, this mechanism ensures that the selected resource blocks are safely separated in the frequency domain. The interference only affects one resource block, while the remaining discretely distributed resource blocks can still successfully transmit data and recover the original information using channel coding redundancy, thereby improving the transmission success rate of high-priority services in harsh electromagnetic environments.

[0052] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the mask mapping unit 1533 is used to perform comb-shaped mask mapping processing on a time-frequency two-dimensional grid to obtain a resource block allocation matrix for the preferred discrete frequency index set. It should be noted that, given that the previously generated preferred discrete frequency index set is only an abstract logical set, it cannot be directly recognized by the underlying field-programmable gate array or digital signal processor as an executable physical layer scheduling instruction. Furthermore, the block masking method commonly used in existing solutions easily covers continuous frequency bands, leading to the destruction of critical system bit regions in background service transmission blocks, thereby causing decoding failures and retransmissions of enhanced mobile broadband services such as on-site monitoring video. Based on this, the technical solution of this application further performs comb-shaped mask mapping processing on a time-frequency two-dimensional grid to obtain a resource block allocation matrix for the preferred discrete frequency index set. Through the above processing, discontinuous preferred frequency points can be accurately mapped to the physical layer time-frequency resource grid, generating punched masks with comb-like or scattered features, thereby achieving precise resource reservation at the physical layer. This effectively avoids strong electromagnetic interference frequency bands and preserves the key system bits of background services to the maximum extent, realizing global optimization of power private network communication performance.

[0053] More specifically, in a concrete example of this application, a discrete comb-shaped punched mask mapping generation operation is performed. Unlike the block mask of the original scheme, this step maps the selected discrete index set onto a time-frequency two-dimensional grid to generate a punched mask with comb-shaped or scattered features. The specific calculation formula in this process is as follows: in, A matrix (i.e., a mask matrix) is assigned to the generated discrete resource blocks. This matrix defines the final resource occupancy state at the physical layer. and These are frequency domain and time domain indexes, used to locate specific resource units. For set The first in The preferred frequency index represents the frequency points with the best channel quality and the least impact on background services, which were previously selected. It is a discrete Dirac function if and only if The value is 1 if the condition is met, and 0 otherwise; this is used for precise selection of the set. The frequency points are selected to ensure that the mask only takes effect at the preferred frequency points. This is a time-domain indicator function, indicating that the mask covers the entire time slot or a specific micro-time slot of the current schedule, used to control the duration of reserved resources in the time dimension. The formula mathematically describes how to transform a discontinuous set of indices into a two-dimensional binary matrix for underlying FPGA or DSP processing. Taking a substation inspection scenario as an example, when a relay protection device triggers an emergency trip command, the system generates a comb-shaped mask according to this formula. This mask exhibits a discontinuous comb-like distribution in the frequency domain, accurately avoiding specific frequency pulse noise caused by disconnector operation. Simultaneously, the gaps between the comb teeth allow critical system bits in the high-definition video stream transmitted back by the background inspection drone to pass through, ensuring that the video image does not exhibit mosaic or interruption during command transmission, achieving a dual guarantee of highly reliable control command transmission and video service continuity.

[0054] In the aforementioned power communication private network management system 100 based on 5G slicing technology, the scheduling execution module 160 is used to control the base station scheduler to perform slice-level resource allocation based on the resource block allocation matrix, prioritizing the transmission of high-priority service data packets on reserved time-frequency resource blocks. It should be noted that, given that conventional base station scheduling mechanisms often employ a retrospective polling strategy based on buffer states, when high-priority relay protection instructions collide with high-volume background video services, they are often forced to queue due to the lack of an absolute priority channel at the physical layer, resulting in end-to-end latency failing to meet the millisecond-level hard real-time constraints of power control services. Based on this, the technical solution of this application further controls the base station scheduler to perform slice-level resource allocation based on the resource block allocation matrix, prioritizing the transmission of high-priority service data packets on reserved time-frequency resource blocks. According to the mask state in the matrix, the transmission queues in the base station buffer pool are split, forcibly filling the key instructions in the high-priority sub-queues into the reserved positions marked by the mask, while filling the background data in the low-priority sub-queues into non-reserved positions. Through the above processing, a green transmission channel independent of background traffic can be effectively built at the physical layer, ensuring that even in the case of network congestion, sudden power control messages can be transmitted instantly with zero waiting time, thereby strictly guaranteeing the deterministic low latency and high reliability of production control slice services.

[0055] More specifically, in a specific example of this application, the scheduling execution module 160 includes: a splitting and assembly unit, used to split the queue of data packets to be transmitted in the base station buffer pool according to the mask state in the resource block allocation matrix to obtain the assembled transport block stream, wherein data packets in the high-priority sub-queue are preferentially filled in the mask reserved positions, and data packets in the low-priority sub-queue are filled in the non-reserved positions; and a signal transformation unit, used to perform channel coding, scrambling modulation and orthogonal frequency division multiplexing transformation on the assembled transport block stream to obtain the radio frequency transmission signal.

[0056] Accordingly, the splitting and assembly unit is used to split the queue of data packets to be transmitted in the base station buffer pool according to the mask state in the resource block allocation matrix to obtain the assembled transport block stream. Specifically, data packets from high-priority sub-queues are preferentially filled in the mask-reserved positions, while data packets from low-priority sub-queues are filled in the non-reserved positions. It should be noted that since the resource block allocation matrix only defines the reservation state of time-frequency resources at the logical level, if the media access control layer lacks a matching mandatory mapping mechanism, high-priority service data packets to be transmitted may still be queued after ordinary services due to the lag in the scheduling algorithm, resulting in idle reserved channels at the physical layer and increased queuing latency at the logical layer. Based on this, the technical solution of this application further splits the queue of data packets to be transmitted in the base station buffer pool according to the mask state in the resource block allocation matrix to obtain the assembled transport block stream. Specifically, data packets from high-priority sub-queues are preferentially filled in the mask-reserved positions, while data packets from low-priority sub-queues are filled in the non-reserved positions. Through the above processing, the predictive scheduling strategy generated at the upper layer can be effectively transformed into the bit stream assembly action at the lower layer, ensuring that high-priority power control commands can be correctly entered into the reserved time and frequency resource blocks, achieving zero-wait transmission at the physical layer, while maintaining normal throughput of background services in non-reserved areas.

[0057] More specifically, in a concrete example of this application, a queue splitting operation is first performed, scanning the queue of data packets to be transmitted in the base station buffer pool and separating them into high-priority sub-queues and low-priority sub-queues according to the service level agreement identifier of the service. Next, a resource mapping operation based on mask priority is performed, traversing each time-frequency resource element in the resource block allocation matrix and determining the source of the filling content based on the mask state at that position. During this process, the mapping logic of the resource block content follows a specific functional relationship: in, Represents resource block The final data content it carries This represents the mask value in the input resource block allocation matrix, where 1 represents the reserved state and 0 represents the normal state. This indicates a high-priority subqueue, such as a relay protection command queue. This indicates a low-priority subqueue, such as the video inspection data queue. This indicates the operation of retrieving a data packet from the head of the queue. These represent padding bits, used to maintain the integrity of the time slot structure or prevent interference to neighboring cells in false alarm scenarios where predictions trigger reservations but no high-priority packets are actually sent. Finally, a transport block encapsulation operation is performed to assemble the mapped and padded data stream into physical layer transport blocks for subsequent processing by the channel coding module.

[0058] Accordingly, the signal transformation unit is used to perform channel coding, scrambling modulation, and orthogonal frequency division multiplexing transformation on the assembled transport block stream to obtain the radio frequency transmission signal. It should be noted that since the assembled transport block stream is essentially still a binary bit sequence of the logic layer, if it is directly transmitted in a complex wireless channel, it is highly susceptible to multipath fading, additive white Gaussian noise, and pulse interference specific to substations, resulting in bit errors. Furthermore, it cannot adapt to the frequency domain transmission characteristics of the physical air interface. Based on this, the technical solution of this application further performs channel coding, scrambling modulation, and orthogonal frequency division multiplexing transformation on the assembled transport block stream to obtain the radio frequency transmission signal. Through the above processing, forward error correction redundancy can be effectively added to high-priority power control data, mapping the logic data into an orthogonal time-frequency waveform with strong anti-interference capability, thereby ensuring reliable bit-level transmission and physical coverage in harsh electromagnetic environments.

[0059] More specifically, in a concrete example of this application, channel coding is first performed. Different coding strategies are adopted for different channel types in the transport block stream. For data channels carrying power service data, low-density parity-check codes are used to obtain error correction performance under high throughput. For control channels carrying scheduling instructions, polar codes are used to ensure the reliability of short code block transmission. Next, scrambling and modulation operations are performed. The encoded bit stream is scrambled using a pseudo-random sequence generated by the physical cell identifier to randomize inter-cell interference. Based on the channel quality results fed back by the link adaptive algorithm, the scrambled bits are mapped to complex modulation symbols. For video services in high signal-to-noise ratio environments, high-order orthogonal amplitude modulation (such as 256QAM) is used, while for control services in low signal-to-noise ratio environments, orthogonal phase-shift keying modulation is used. Subsequently, orthogonal frequency division multiplexing (OFDM) generation is performed. The frequency domain complex symbols are converted into time domain waveforms using inverse fast Fourier transform, and a cyclic prefix is ​​added to the front of the symbols to eliminate delay spread and inter-symbol interference caused by multipath propagation. Finally, the digital-to-analog conversion and radio frequency transmission operations are performed. The digital baseband waveform is converted into an analog signal by the digital-to-analog converter and loaded onto the carrier frequency. The electromagnetic wave signal is then transmitted to the power terminal and public network users through the antenna array.

[0060] In summary, the power communication private network management system based on 5G slicing technology according to the embodiments of this application is explained. It constructs a cross-domain integrated predictive scheduling mechanism, utilizing the heterogeneous data timing characteristics of the power grid and network sides to predict the burst probability of high-priority services in advance. Based on this, combining the power impulse noise spectrum and the background service coding mapping table, a comprehensive utility evaluation, including power grid environment awareness and coding impairment costs, is performed on physical resource blocks. A discrete set of anti-interference frequency points in the frequency domain is selected and mapped to a comb-shaped punch mask. Then, the base station scheduler is controlled to establish dedicated transmission channels for high-priority services at the physical layer based on the mask state, achieving a logical leap from passive queuing to active awareness. This not only effectively avoids strong electromagnetic interference frequency bands and ensures zero-wait transmission of power control commands, but also reduces the decoding impact on background enhanced mobile broadband services by protecting critical system bits, achieving panoramic optimization of heterogeneous services under limited spectrum resources.

Claims

1. A power communication private network management system based on 5G slicing technology, characterized in that, include: The data acquisition module is used to synchronously acquire raw grid-side data containing instantaneous values ​​of three-phase voltage and current, as well as raw network-side data containing slice buffer queue length and channel quality indication, in real time through the edge computing gateway. The data alignment module is used to align the timestamps of the raw data from the power grid side and the raw data from the network side to obtain a joint raw dataset. The feature fusion module is used to normalize the joint original dataset and extract high-frequency transient component features, and then concatenate the extracted feature vector with the network state vector in multiple dimensions to obtain a fusion state tensor that represents the comprehensive panoramic state of the power grid and network at the current moment. The prediction and extrapolation module is used to predict and extrapolate the burst probability of high-optimal services on the fusion state tensor to obtain the burst probability value. The resource decision module is used to determine the resource block allocation matrix to guide the medium access control layer in resource scheduling based on the burst probability value and the channel quality data in the joint original dataset. The scheduling execution module is used to control the base station scheduler to perform slice-level resource allocation based on the resource block allocation matrix, and to prioritize the transmission of high-priority service data packets on the reserved time and frequency resource blocks.

2. The power communication private network management system based on 5G slicing technology according to claim 1, characterized in that, The data alignment module includes: The asynchronous buffer unit is used to asynchronously buffer and unpack the original data from the power grid side and the network side based on a dual-channel ring buffer mechanism to obtain heterogeneous buffered data pairs. The time base unification unit is used to unify the time base and extract the time scale of heterogeneous buffered data pairs to obtain a standardized time-series data stream with a unified time dimension. The resampling alignment unit is used to perform sliding window-based resampling and temporal alignment on the normalized time-series data stream to obtain a joint original dataset.

3. The power communication private network management system based on 5G slicing technology according to claim 1, characterized in that, The feature fusion module includes: The normalization processing unit is used to normalize the electrical data and network data in the joint original dataset to obtain normalized data pairs; The feature extraction unit is used to extract high-frequency transient and trend features from normalized data pairs in a dual-track manner to obtain a heterogeneous feature vector set. The sliding window stacking unit is used to slide window stack the multimodal features in the heterogeneous feature vector set along the time step dimension to obtain the fused state tensor.

4. The power communication private network management system based on 5G slicing technology according to claim 1, characterized in that, The prediction and inference module includes: The deep extraction unit is used to perform temporal feature deep extraction on the fused state tensor based on a recurrent neural network to obtain the hidden layer feature sequence. Attention-focusing unit is used to perform attention-weighted focusing on the hidden layer feature sequence at key time steps to obtain a contextual semantic vector focused on fault symptoms; The probability mapping unit is used to perform nonlinear probability mapping and confidence output on the context semantic vector to obtain the burst probability value.

5. The power communication private network management system based on 5G slicing technology according to claim 1, characterized in that, The resource decision-making module includes: The threshold comparison unit is used to perform a hysteresis comparison between the burst probability value and the preset high water level threshold and low water level threshold to obtain a preemption trigger flag indicating whether to perform resource reservation. The resource quantity calculation unit is used to find the corresponding spectral efficiency based on the channel quality indication in the joint original dataset when the preemption trigger flag is valid, and to calculate the required frequency domain resources for transmission in combination with the preset target payload size to obtain the required resource block count. The matrix determination unit is used to determine the resource block allocation matrix based on the required resource block count.

6. The power communication private network management system based on 5G slicing technology according to claim 1, characterized in that, The scheduling execution module includes: The splitting and assembly unit is used to split the queue of data packets to be transmitted in the base station buffer pool according to the mask status in the resource block allocation matrix to obtain the assembled transport block stream. In this case, data packets in the high-priority sub-queue are filled first in the mask reserved positions, and data packets in the low-priority sub-queue are filled in the non-reserved positions. The signal transformation unit is used to perform channel coding, scrambling modulation, and orthogonal frequency division multiplexing transformation on the assembled transport block stream to obtain the radio frequency transmission signal.

7. The power communication private network management system based on 5G slicing technology according to claim 5, characterized in that, The matrix determination unit includes: The region selection unit is used to perform region selection and mask marking in the time-frequency two-dimensional resource grid according to the required resource block count; The state setting unit is used to set the elements in the selected area to the reserved state in order to obtain the resource block allocation matrix.

8. The power communication private network management system based on 5G slicing technology according to claim 5, characterized in that, The matrix determination unit includes: The utility evaluation unit is used to perform a comprehensive utility evaluation on each physical resource block within the full bandwidth based on the subcarrier signal-to-noise ratio data, the preset power impulse noise spectrum and the background service coding mapping table in the joint original dataset, so as to obtain a resource utility score vector. The optimal decision unit is used to perform a greedy optimal decision with frequency domain discrete constraints on the resource utility score vector and the required number of physical resource blocks to obtain the optimal discrete frequency index set. The mask mapping unit is used to perform comb-shaped mask mapping processing on the preferred discrete frequency index set on a time-frequency two-dimensional grid to obtain the resource block allocation matrix.

Images