A multi-mode transmission scheduling method and system for a low-voltage area intelligent acquisition terminal

Abstract

This invention discloses a multi-mode transmission scheduling method and system for intelligent data acquisition terminals in low-voltage distribution areas. The method extracts time-series feature vectors of electricity consumption behavior from historical operational data of the distribution area's data acquisition terminals, trains a recurrent neural network channel quality prediction model based on a time attention mechanism, and outputs a sequence of channel quality levels within a future time window. It defines transmission demand vectors for each service type and uses a two-stage decision-making mechanism to generate a service-to-channel allocation scheme. In a power outage message storm scenario, power outage events are classified into three modes based on their temporal spread rate and spatial spread range, and communication resources are partitioned and reconstructed. Online incremental updates of the channel quality prediction model are triggered by collecting actual transmission results. This invention achieves predictive proactive switching of communication modes, prioritizing delay-sensitive services and ensuring complete and timely reporting of power outage range information even with minimal available channel resources.

Description

Technical Field

[0001] This invention belongs to the field of intelligent electricity consumption information collection technology in power systems, specifically relating to a multi-mode transmission scheduling method and system for intelligent collection terminals in low-voltage distribution areas. Background Technology

[0002] The low-voltage distribution transformer intelligent data acquisition terminal is the core terminal device of the power consumption information acquisition system. Deployed on the distribution transformer side of the low-voltage distribution area, it undertakes the task of collecting and uploading electricity consumption data from all electricity meters within its jurisdiction. With the continuous development of communication technology, mainstream acquisition terminals now possess multi-mode communication hardware capabilities. They can upload collected data to the main station system through various communication modes such as HPLC high-speed power line carrier, low-power wireless, GPRS, 4G, 5G, and fiber optics. Some systems have also deployed a dual-mode fusion networking scheme of HPLC and HRF radio frequency. The above-mentioned multi-mode communication capabilities provide a physical basis for improving the reliability of distribution transformer data acquisition. However, how to reasonably and dynamically schedule multiple communication modes has become a key issue restricting the operational efficiency of the acquisition system.

[0003] However, current communication mode scheduling strategies for multi-mode acquisition terminals generally employ a passive switching logic based on a fixed threshold using a single physical indicator. This means that communication mode switching is only triggered when the signal strength falls below a preset threshold or the packet loss rate exceeds a preset threshold. This passive response method inherently suffers from scheduling lag. The power line channel in low-voltage distribution areas is directly affected by the load equipment on the user side. Switching transients of this equipment generate broadband impulse noise and harmonic interference on the power lines, thus degrading the communication quality of the HPLC channel. The usage behavior of these load devices exhibits strong temporal regularity; for example, noise levels during peak electricity consumption periods are significantly higher than during off-peak periods, and there are systematic differences in channel noise characteristics between holidays and weekdays. Because the changes in power line channel quality have predictable time-varying patterns, each decision in the passive switching strategy occurs after data loss has already occurred within the channel degradation window. It is impossible to proactively complete the preparatory switching of communication modes before channel degradation occurs, which has a particularly significant impact on delay-sensitive services such as power outage reporting. Currently, there is no systematic research or application of methods to establish causal prediction models for channel quality based on the temporal characteristics of power consumption behavior in distribution substations, thereby enabling predictive and proactive switching of communication modes.

[0004] Furthermore, the types of concurrent services operating within the distribution area are diverse, ranging from power outage reporting, which is extremely sensitive to latency, to periodic meter reading, which has high reliability requirements but relatively low latency tolerance, and power quality waveform uploading, which requires a certain amount of transmission bandwidth. These different service types exhibit significant differences in transmission requirements, such as maximum permissible latency, minimum reliability requirements, and estimated data volume. However, existing scheduling strategies do not incorporate these differentiated transmission requirements into channel allocation decisions, treating all services indiscriminately in terms of communication resource allocation. This approach is not prominent when communication resources are abundant, but when channel quality deteriorates and available resources become scarce, latency-sensitive services and low-priority services will compete for the same channel, compromising the transmission latency of critical services. The aforementioned problem of decoupling service requirements from channel state scheduling remains unresolved in the existing scheduling design of multi-mode acquisition terminals.

[0005] The most severe scenario is a sudden large-scale power outage in a distribution area. When a large-scale power outage occurs, a large number of data acquisition nodes within the area detect the outage almost simultaneously and trigger emergency reporting messages, creating a high-concurrency competition for communication resources. In this scenario, the HPLC carrier channel partially or completely fails due to the power interruption, forcing a large number of terminals to flood the wireless channel, resulting in severe congestion and loss or significant delay of critical outage event messages. The power outage event itself causes both carrier channel failure and a large number of outage messages to be generated, rendering traditional scheduling methods based on historical channel data completely ineffective in this extreme scenario. The master station system needs to obtain accurate outage range information as soon as possible after the outage to guide emergency repair scheduling. Currently, for this extreme resource conflict state with the highest service priority and a sharp contraction of available channels, there is no specific outage pattern recognition and partitioned communication resource reconstruction scheme in existing technologies, which cannot effectively cope with the impact of outage message storms on critical service transmissions.

[0006] Regarding adaptive model updates, after the data acquisition terminals are deployed on-site, the power consumption structure and load composition of the distribution area continue to evolve over time, potentially causing a decrease in the prediction accuracy of the initially trained channel quality prediction model. Online incremental updates are one approach to address this issue, but direct incremental updates, while introducing new data, can interfere with the model's learned historical knowledge, resulting in a catastrophic forgetting problem and a degradation of the model's predictive ability in older scenarios. Existing intelligent scheduling solutions for distribution area data acquisition terminals lack specific methods for achieving online adaptive adjustment of model parameters while preserving historical predictive capabilities. Summary of the Invention

[0007] To address the technical problems of existing low-voltage distribution area multi-mode acquisition terminals, such as delayed passive switching of communication modes, decoupling of service transmission requirements and channel scheduling, and lack of effective scheduling mechanisms in power outage message storm scenarios, this invention provides a multi-mode transmission scheduling method and system for intelligent acquisition terminals in low-voltage distribution areas. This method constructs a channel quality prediction model based on the time-series characteristics of distribution area electricity consumption behavior, establishes a two-dimensional joint scheduling framework for service requirements and channel status, and designs a rapid reconstruction mechanism for emergency communication resources in power outage scenarios, thereby achieving predictive proactive switching of communication modes and differentiated transmission scheduling.

[0008] In one aspect, the present invention provides a multi-mode transmission scheduling method for a low-voltage distribution area intelligent acquisition terminal, comprising the following steps.

[0009] Step S1: Extract the time-series feature vector of electricity consumption behavior from the historical operation data of the distribution area acquisition terminal. The time-series feature vector of electricity consumption behavior includes the time-segmented statistical value of the total active power of the distribution area, the start-stop event count sequence of the load equipment, the power factor change rate sequence, and the type identifier of the day. After aligning the time-series feature vector of electricity consumption behavior with the historical quality indicators of each communication channel during the same period, train a channel quality prediction model based on a recurrent neural network with a time attention mechanism, and output the quality level sequence of each communication channel within a future preset time window.

[0010] Specifically, the time-segmented statistical value of the total active power of the distribution area is a characteristic value obtained by dividing the day into multiple time periods and statistically analyzing the mean, peak value, and standard deviation of the total active power of the distribution area within each time period. The start-stop event count sequence of the load equipment is a time series formed by statistically analyzing the cumulative number of switching actions of various types of load equipment in the distribution area within each time period. The power factor change rate sequence reflects the temporal characteristics of reactive load fluctuations in the distribution area. The day type identifier is used to distinguish between weekdays, weekends, and statutory holidays to capture the systematic differences in electricity consumption behavior under different date types.

[0011] Furthermore, the historical quality indicators include the signal-to-noise ratio and frame error rate of the HPLC channel, and the received signal strength indicator and round-trip time of the wireless channel. The time-series feature vector of electricity consumption behavior is aligned with the aforementioned historical quality indicators according to timestamps to form an input-label paired training dataset. The aligned time-series feature vector of electricity consumption behavior is used as the model input, and the quality level of each channel at the corresponding time moment is used as the prediction label to train the channel quality prediction model.

[0012] Furthermore, the encoder of the channel quality prediction model employs a gated recurrent unit to perform temporal encoding on the electricity consumption behavior time-series feature vector. The decoder introduces a temporal attention layer to weighted aggregate the hidden states of the encoder at each time step. This temporal attention layer uses an additive attention calculation method, and its attention weights are calculated by concatenating the encoder's hidden state and the decoder's current hidden state, mapping them through a fully connected layer, and then calculating them using a normalized exponential function. Through the temporal attention layer, the model can adaptively focus on the historical time-series segments that have the greatest impact on the current prediction window, extracting the periodic patterns and trend features of electricity consumption behavior.

[0013] Preferably, the preset time window is 15 to 60 minutes, and the quality level sequence is divided into four discrete levels: excellent, good, medium, and poor. These four discrete levels correspond to the maximum number of service types the channel can carry. An excellent level indicates that the channel can carry all service types, while a poor level indicates that the channel can only carry services with low reliability requirements or is unavailable.

[0014] Furthermore, the channel quality prediction model employs a weighted cross-entropy loss function during training, assigning a higher penalty weight to prediction errors occurring when channel quality deteriorates compared to those occurring when it deteriorates. This asymmetric penalty design encourages the model to conservatively estimate channel quality during prediction, reducing the risk of service transmission failures due to missed channel degradation events. Preferably, the penalty weight for prediction errors occurring when channel quality deteriorates is set to 2 to 5 times the penalty weight for those occurring when it deteriorates.

[0015] Step S2: Define a transmission demand vector for each service type. The transmission demand vector includes the maximum allowable delay, the minimum reliability requirement, and the estimated data volume. Based on the quality level sequence and the transmission demand vector, a two-stage decision-making mechanism is used to generate a service-to-channel allocation scheme.

[0016] Specifically, the data acquisition terminals in the distribution area handle services including power outage event reporting, periodic meter reading, and power quality waveform uploading. The maximum permissible latency for power outage event reporting is set at 3 seconds, with a minimum reliability requirement of 99.9%; the maximum permissible latency for periodic meter reading is set at 300 seconds, with a minimum reliability requirement of 99%; and the maximum permissible latency for power quality waveform uploading is set at 60 seconds, with a minimum reliability requirement of 95%.

[0017] Furthermore, in the two-stage decision-making mechanism, the first stage is the constraint satisfaction stage, which traverses the queue of services to be transmitted from high to low priority, and selects a set of candidate channels for each service whose predicted quality level meets the transmission demand vector constraints; the second stage is the global optimization stage, which aims to minimize the weighted transmission completion time of all services and solves the allocation scheme using a dynamic programming method. The weight of each service is proportional to the reciprocal of its maximum allowable delay. By assigning higher weights to delay-sensitive services, they are given priority in resource contention.

[0018] Preferably, the two-stage decision-making mechanism also introduces a channel handover cost evaluation mechanism. When the decision requires a service to switch from the current channel to another channel, the handover cost is evaluated using a maintained historical statistics table of handover delays between channels. The reduction in the expected transmission completion time of the service after the handover is taken as the handover benefit. The handover operation is only performed when the handover benefit is greater than a preset safety margin multiple of the handover cost, preferably 1.5. This mechanism avoids the introduction of additional delay overhead due to frequent channel handovers and ensures the stability of scheduling decisions.

[0019] Step S3: When the number of power outage event reporting requests exceeds the storm threshold within the preset time window, the power outage events are classified into three modes based on the time diffusion rate and spatial diffusion range of the power outage events: local power outage, branch power outage, and whole-area power outage. The communication resources are then partitioned and reconstructed according to the power outage mode.

[0020] Specifically, the rule for classifying power outage events into three modes based on the time diffusion rate and spatial diffusion range of the power outage event is as follows: when the spatial diffusion range is limited to a single branch line and the time diffusion rate is lower than a first threshold, it is determined to be a local power outage; when the spatial diffusion range covers multiple branch lines but not all of them and the time diffusion rate does not exceed a second threshold, it is determined to be a branch power outage; when the spatial diffusion range covers the entire transformer area or the time diffusion rate exceeds the second threshold, it is determined to be a full transformer area power outage; wherein the time diffusion rate is defined as the maximum value of the number of newly reported power outage nodes in each unit time slot within a preset time window, and the spatial diffusion range is defined as the coverage range of the connected subgraph of the transformer area topology map of the reported power outage nodes.

[0021] Furthermore, for the partial power outage mode, all reporting services from nodes within the outage area are scheduled to the wireless channel, while the non-outage area maintains its regular scheduling. For the branch power outage mode, the wireless channel capacity is equally divided according to the number of outage branches, and an independent reporting time slot window is allocated to each outage branch. The reporting order of outage reporting nodes within each branch is determined by the hash value of the node address. For the full-area power outage mode, all wireless channel capacity is allocated to the outage reporting service. The outage reporting message is simplified and encoded, retaining only three fields: node identifier, outage timestamp, and last voltage sample value. Nodes with remaining power supply act as proxies to collect outage information from their downstream outage nodes and report it uniformly in the form of aggregated messages. Through the aggregated reporting mechanism, the number of access contention times for the wireless channel is reduced, and the length of a single message is compressed to the minimum message length, ensuring complete reporting of outage range information with minimal available channel resources.

[0022] Step S4: Collect the actual transmission results of each service and compare them with the expected results during scheduling decisions. When the channel quality level prediction accuracy is lower than the preset lower limit, trigger the online incremental update of the channel quality prediction model.

[0023] Specifically, at the end of each scheduling cycle, the acquisition terminal records the actual transmission completion time, actual channel usage, and transmission success rate of each service. These actual results are then compared with the expected results at the time of scheduling decision-making to calculate a prediction deviation index. The prediction deviation index includes the accuracy of channel quality level prediction and the prediction error rate of service transmission completion time.

[0024] Furthermore, the online incremental update employs a resilient weight consolidation method. When updating model parameters, regularization constraints are applied to learned important parameters, and the running data within the most recent preset number of cycles is used as incremental training samples to fine-tune the parameters of the channel quality prediction model. The resilient weight consolidation method evaluates the importance of each parameter to the learned knowledge using the Fisher information matrix, and imposes constraints on the update magnitude of important parameters to prevent catastrophic forgetting of historical knowledge during the incremental update process. Preferably, the preset number of cycles is the most recent 7 to 30 scheduling cycles, and the learning rate for incremental training is set to one-tenth to one-hundredth of the initial training learning rate.

[0025] Furthermore, the closed-loop feedback also includes adaptive adjustment of the power outage pattern recognition rule parameters in step S3. After each power outage event is processed, the storm threshold, the first threshold, and the second threshold are corrected based on the actual power outage range information fed back by the main station, so as to continuously improve the classification accuracy of the power outage pattern.

[0026] In another aspect, the present invention provides a multi-mode transmission scheduling system for a low-voltage distribution area intelligent data acquisition terminal, comprising: an electricity consumption behavior feature extraction module, used to extract a time-series feature vector of electricity consumption behavior from the historical operating data of the distribution area data acquisition terminal, wherein the time-series feature vector of electricity consumption behavior includes time-segmented statistical values ​​of the total active power of the distribution area, a sequence of start-stop events of load equipment, a sequence of power factor change rate, and a daily type identifier; a channel quality prediction module, used to time-align the time-series feature vector of electricity consumption behavior with the historical quality indicators of each communication channel during the same period, and output a sequence of quality levels of each communication channel within a future preset time window through a channel quality prediction model based on a recurrent neural network with a time attention mechanism; and a service demand modeling module, used to define a transmission demand vector for each service type, wherein the transmission demand vector... This includes the maximum permissible latency, minimum reliability requirements, and estimated data volume; a joint scheduling decision module, used to generate a service-to-channel allocation scheme based on the quality level sequence output by the channel quality prediction module and the transmission demand vector using a two-stage decision mechanism; a power outage emergency reconstruction module, used to classify power outage events into three modes—local power outage, branch power outage, and full-area power outage—based on the time spread rate and spatial spread range of the power outage events when the number of power outage event reporting requests exceeds the storm threshold within a preset time window, and to perform partitioned reconstruction of communication resources according to the power outage mode; and a closed-loop feedback update module, used to collect the actual transmission results of each service, compare them with the expected results during scheduling decisions, and trigger online incremental updates of the channel quality prediction model when the channel quality level prediction accuracy is lower than a preset lower limit.

[0027] The electricity consumption behavior feature extraction module reads raw data from the historical operation database stored in the transformer area acquisition terminal according to time windows. After data cleaning and outlier removal, it calculates statistical features of each dimension according to a preset time period division method, forming an electricity consumption behavior time-series feature vector, which is then passed to the channel quality prediction module. The channel quality prediction module loads the trained channel quality prediction model, receives the electricity consumption behavior time-series feature vector as input, outputs the predicted quality level sequence of each channel within the preset time window, and passes the prediction results to the joint scheduling decision module. The service requirement modeling module maintains the transmission requirement vector configuration table for each service type and provides the joint scheduling decision module with the transmission requirement constraints of each service according to the current queue of services to be transmitted. The joint scheduling decision module executes a two-stage decision mechanism, integrates the channel quality prediction results and service transmission requirements, outputs the service-to-channel allocation scheme, and sends it to the communication interface for execution. The power outage emergency reconstruction module continuously monitors the arrival rate of power outage event reporting requests. When a power outage message storm is detected, it takes over the scheduling control, performs power outage pattern recognition and communication resource partition reconstruction, and returns the scheduling control to the joint scheduling decision module after the power outage event is processed. The closed-loop feedback update module collects the scheduling execution results after each scheduling cycle, calculates the prediction deviation index, and drives the online incremental update of the channel quality prediction model and the adaptive correction of the power outage mode identification parameters when the triggering conditions are met.

[0028] This invention constructs a channel quality prediction model based on the temporal characteristics of electricity consumption behavior. It uses the temporal patterns of electricity consumption behavior in distribution areas as a leading indicator of channel quality changes, achieving predictive awareness of channel status. This transforms communication mode switching from passive response to proactive prediction, effectively avoiding data loss during channel degradation windows. Furthermore, this invention establishes a two-stage joint scheduling decision mechanism based on service awareness. It incorporates the differentiated transmission requirements of different service types into scheduling decisions, combining constraint satisfaction and global optimization to achieve a reasonable allocation of communication resources among various services. This prioritizes latency-sensitive services while maintaining overall transmission efficiency. Finally, this invention designs a rapid emergency communication resource reconstruction mechanism for power outage scenarios. For the extreme scenario of power outage message storms, it establishes a complete emergency scheduling process from power outage pattern identification to resource partitioning and reconstruction. Through hierarchical compression reporting and message aggregation mechanisms, it ensures complete and timely reporting of power outage range information even with minimal available channel resources. Finally, this invention introduces a closed-loop feedback and online incremental update mechanism, enabling the channel quality prediction model and scheduling parameters to continuously and adaptively optimize with changes in the distribution area's operating environment, ensuring the stability and accuracy of the scheduling system in long-term operation. Attached Figure Description

[0029] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] Figure 1 This is a schematic diagram of the overall process of the multi-mode transmission scheduling method for the intelligent acquisition terminal in low-voltage distribution areas provided in this embodiment of the invention.

[0031] Figure 2 This is a schematic diagram of the channel quality prediction model based on a recurrent neural network with a time attention mechanism provided in an embodiment of the present invention.

[0032] Figure 3 This is a schematic diagram of the processing flow of the emergency communication resource reconstruction mechanism in a power outage scenario provided by an embodiment of the present invention.

[0033] Figure 4 This is a block diagram of the module structure of the multi-mode transmission scheduling system of the low-voltage distribution area intelligent acquisition terminal provided in this embodiment of the invention.

[0034] Figure 5 This is a timing comparison diagram of multi-mode channel quality prediction and service scheduling provided in an embodiment of the present invention.

[0035] Figure 6 This is a constraint-optimization space visualization diagram of the two-stage joint scheduling decision-making mechanism provided in the embodiments of the present invention.

[0036] Figure 7 This is a physical deployment scenario diagram of the low-voltage distribution area multi-mode communication acquisition terminal provided in an embodiment of the present invention. In the diagram, 101 is the distribution area transformer, 102 is the acquisition terminal, 103 is the antenna, 104 is the main line, 105 is the power pole, 106 is the branch line, 107 is the acquisition node, 108 is the industrial plant, 109 is the industrial motor, 110 is the residential building, and 111 is power line noise.

[0037] Figure 8 This is a composite graph comparing the performance of various ablation experimental schemes provided in the embodiments of the present invention.

[0038] Figure 9 This is a graph showing the relationship between the information reporting completeness rate and time in a power outage scenario, as provided in an embodiment of the present invention.

[0039] Figure 10 This is a heatmap showing the training convergence and long-term operating performance of the channel quality prediction model provided in this embodiment of the invention. Detailed Implementation

[0040] To make the objectives and technical solutions of this invention clearer, the embodiments of this invention will be further described in detail below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of this invention and should not be construed as limiting the scope of protection of this invention.

[0041] Example 1

[0042] like Figure 7As shown, the physical deployment layout of the low-voltage distribution area acquisition terminal (i.e., multi-mode communication acquisition terminal, hereinafter referred to as the acquisition terminal) in the actual distribution area is as follows. The distribution area transformer 101 is located at the center of the upper part of the scene. Its shape is presented as an oil tank body, heat dissipation fins on both sides, high-voltage bushing at the top, and low-voltage terminal block at the bottom. The acquisition terminal 102 is installed on the lower side of the distribution area transformer 101. The terminal box is equipped with a display panel, status indicator lights, communication interface, and heat dissipation grid. The antenna 103 is configured on the top of the box for low-power wireless channel communication. A main line 104 is led out from the acquisition terminal 102, and power poles 105 are erected along the line. The main line 104 extends to both sides as four branch lines 106. The left side is an industrial plant area, where industrial plants 108 and industrial motors 109 are deployed. The power line noise 111 generated by the industrial motor 109 during operation propagates along the cable, showing its interference to the HPLC high-speed power line carrier channel in the form of ripples. The right side is a residential area, where multiple residential buildings 110 are distributed. Each branch line 106 has a data acquisition node 107 distributed at its end. The data acquisition node 107 is presented in the form of an LCD display window, a metering turntable, and a terminal block. The communication channel includes two physical paths: the HPLC high-speed power line carrier channel transmits along the main line 104 and the branch line 106, and the low-power wireless channel transmits via antenna 103 to the remote data acquisition node 107 using radio waves. The two channels complement each other, providing a physical basis for multi-mode transmission scheduling.

[0043] like Figure 1 As shown, this embodiment provides a multi-mode transmission scheduling method for a low-voltage distribution transformer area intelligent acquisition terminal. This method is applied to a multi-mode communication acquisition terminal deployed on the distribution transformer area side of a low-voltage transformer. The acquisition terminal has two communication modes: HPLC high-speed power line carrier channel and low-power wireless channel. In this embodiment, the number of acquisition nodes under the jurisdiction of the distribution transformer area is 256, and the scheduling cycle of the acquisition terminal is set to 15 minutes. The method includes steps S1 to S4, which are described in detail below.

[0044] Step S1: Extract the time-series feature vector of electricity consumption behavior from the historical operation data of the collection terminal in the distribution area. After aligning the time-series feature vector of electricity consumption behavior with the historical quality indicators of each communication channel during the same period, train a channel quality prediction model based on a recurrent neural network with a time attention mechanism, and output the quality level sequence of each communication channel within a future preset time window.

[0045] In this embodiment, the historical operating data stored in the distribution area acquisition terminal is first read and preprocessed. The distribution area acquisition terminal continuously records the electrical operating parameters of the distribution area, including the three-phase voltage, three-phase current, active power, reactive power, power factor, etc. of the low-voltage side of the distribution area transformer, with a sampling period of 1 minute. A time-series feature vector of electricity consumption behavior is extracted from the above raw data. This vector consists of feature components in the following four dimensions. The first dimension is the time-segmented statistical value of the total active power of the distribution area. A 24-hour day is divided into 96 15-minute time segments. The mean, peak value, and standard deviation of the sampled total active power of the distribution area within each time segment are calculated to form a feature sub-vector of length 288. The second dimension is the load equipment start-stop event counting sequence. By performing threshold detection on the first-order difference of the total active power of the distribution area, the switching action events of the load equipment are identified. When the absolute value of the active power difference between two adjacent sampling points exceeds 0.5kW, it is recorded as a start-stop event. The cumulative number of start-stop events is counted within each 15-minute time segment to form a counting sequence of length 96. The third dimension is the power factor change rate sequence. The power factor change rate is calculated for each 15-minute period, which is the difference between the power factor at the end of the period and the power factor at the beginning of the period, divided by the period length, forming a change rate sequence of length 95. This sequence reflects the temporal characteristics of reactive load fluctuations within the distribution area. The fourth dimension is the daily type identifier, using unique thermal encoding to represent weekdays, weekends, and public holidays, forming a binary vector of length 3. The feature components of these four dimensions are concatenated to form a complete temporal feature vector of electricity consumption behavior.

[0046] In the data preprocessing stage, the extracted raw feature vectors undergo data cleaning and standardization. For missing values ​​in active power and power factor data caused by communication interruptions or sensor malfunctions, linear interpolation is used to impute them. Outliers that significantly exceed physically reasonable ranges, such as negative active power values ​​or power factors exceeding 1, are marked as invalid and replaced with the mean of adjacent valid values. After cleaning, each dimension of the features is standardized with zero mean and unit variance to eliminate the impact of dimensional differences between different physical quantities on model training.

[0047] The cleaned and standardized time-series feature vector of electricity consumption behavior is time-aligned with the historical quality indicators of each communication channel during the same period. These historical quality indicators include the signal-to-noise ratio (SNR) (in dB) and frame error rate (dimensionless percentage) of the HPLC channel, and the received signal strength index (RSSI, in dBm) and round-trip time (RTT, in milliseconds) of the wireless channel. These quality indicators are aggregated and statistically analyzed in 15-minute time intervals corresponding to the electricity consumption behavior characteristics, and the average value of each indicator within each time interval is taken as the channel quality label for that time interval. Further, based on the comprehensive score of each channel quality indicator, the channel quality is divided into four discrete levels: Excellent, Good, Average, and Poor. Specifically, for the HPLC channel, an SNR greater than 25 dB and a frame error rate less than 1% is marked as Excellent; an SNR between 15 dB and 25 dB and a frame error rate less than 5% is marked as Good; an SNR between 8 dB and 15 dB and a frame error rate less than 15% is marked as Average; and all other cases are marked as Poor. For wireless channels, a channel is classified as Excellent when its RSSI is greater than -70dBm and its RTT is less than 50ms; Good when its RSSI is between -85dBm and -70dBm and its RTT is less than 100ms; Medium when its RSSI is between -95dBm and -85dBm and its RTT is less than 200ms; and Poor in all other cases. Excellent indicates the channel can carry all service types, Good indicates it can carry services except for high-bandwidth waveform uploads, Medium indicates it can only carry low-speed services such as periodic meter readings, and Poor indicates it can only carry services with low reliability requirements or is unavailable. The aligned time-series feature vector of electricity consumption behavior is used as the model input, and the quality level of each channel at the corresponding time moment is used as the prediction label, forming an input-label paired training dataset.

[0048] The following describes the execution flow of the channel quality prediction model during the inference phase. For example... Figure 2As shown, the channel quality prediction model adopts an encoder-decoder architecture. During inference execution, the encoder receives the temporal feature vector sequence of electricity consumption behavior over the most recent T time periods (T is 96 in this embodiment, i.e., the most recent 24 hours) as input. The encoder uses a two-layer gated recurrent unit (GRU) with a hidden state dimension of 128. It performs time-step temporal encoding on the input sequence, outputting a 128-dimensional hidden state vector at each time step. After processing all T time steps, the encoder outputs a hidden state sequence H = {h_1, h_2,..., h_T}. The decoder also adopts a two-layer GRU structure with a hidden state dimension of 128. Its initial hidden state is initialized by the hidden state of the encoder's last time step. The decoder introduces a temporal attention layer at each prediction time step to perform weighted aggregation of the hidden state sequence output by the encoder. The temporal attention layer employs additive attention computation. Specifically, the process is as follows: the hidden state h_i of the encoder at time step i is concatenated with the hidden state s_t of the decoder at the current time step, resulting in a 256-dimensional concatenated vector. This concatenated vector is then mapped through a fully connected layer with an output dimension of 64 and a tanh activation function, followed by another fully connected layer with an output dimension of 1 to obtain a scalar attention score e_i. The attention scores for all time steps are used to calculate the normalized attention weight alpha_i = exp(e_i) / sum(exp(e_j)). The final context vector c_t = sum(alpha_i * h_i) serves as the auxiliary input for the decoder at the current time step, is concatenated with the decoder's own input, and then fed into the GRU unit. Through a time attention layer, the model can adaptively focus on historical time-series segments that have the greatest impact on the current prediction window. For example, when predicting channel quality during peak evening electricity consumption, the attention weights automatically focus on electricity consumption behavior characteristics from the same time period the previous day and the afternoon of the current day, thereby extracting periodic patterns and trend features of electricity consumption behavior. The decoder's output at each prediction time step is mapped to a probability distribution of four quality levels through a fully connected layer, and the level with the highest probability is taken as the predicted quality level for that time step. In this embodiment, the prediction window is set to 30 minutes, meaning the decoder outputs a sequence of channel quality levels for the next two time periods.

[0049] The training phase of the channel quality prediction model is described below. The training dataset is constructed using a sliding window approach, extracting samples from at least 90 days of historical operational data accumulated by the terminals in the distribution area, with a step size of one time period. Each sample contains a time-series feature vector of electricity consumption behavior for 96 consecutive time periods as input, and channel quality levels for the following two time periods as labels. The dataset is divided into training, validation, and test sets in chronological order, with a ratio of 7:1.5:1.5, ensuring that the time windows of the validation and test sets are later than those of the training set to avoid data leakage. A weighted cross-entropy loss function is used during training. This loss function assigns asymmetric penalty weights to prediction errors in different directions: a 3.5-fold penalty weight is given to prediction errors from good to deteriorating channel quality (i.e., actual channel quality deteriorates but the model predicts good), while a 1.0-fold base weight is given to prediction errors from deteriorating to good (i.e., actual channel quality is good but the model predicts deterioration). This asymmetric penalty design makes the model tend to conservatively estimate channel quality during prediction, reducing the risk of service transmission failure due to missed channel degradation events. The optimizer uses the Adam optimizer with an initial learning rate set to 2.3 × 10⁻⁻⁻⁶. 4 The beta_1 parameter is 0.9, and the beta_2 parameter is 0.999. The learning rate decay strategy uses cosine annealing, gradually decreasing the learning rate from its initial value to one-hundredth of its initial value during training. The batch size is set to 64, and the number of training epochs is set to 120. An early stopping mechanism is introduced, terminating training prematurely if the weighted cross-entropy loss on the validation set does not decrease for 15 consecutive epochs. For regularization, Dropout is introduced between GRU layers, with a dropout rate of 0.3. Training hardware uses a server equipped with a single GPU, implemented based on the PyTorch deep learning framework. The model's evaluation metrics on the test set include classification accuracy and weighted F1 score at each quality level. Preferably, the model is considered to have converged when the weighted F1 score reaches 0.85 or higher.

[0050] Step S2: Define a transmission demand vector for each service type. The transmission demand vector includes the maximum allowable delay, the minimum reliability requirement, and the estimated data volume. Based on the quality level sequence and the transmission demand vector, a two-stage decision-making mechanism is used to generate a service-to-channel allocation scheme.

[0051] In this embodiment, the service types and their transmission requirement vectors carried by the distribution area acquisition terminal are defined as follows: The maximum allowable latency for the power outage event reporting service is set to 3 seconds, the minimum reliability requirement is set to 99.9%, and the estimated data volume is 64 bytes per message. This service has the highest transmission priority and must immediately occupy an available channel for transmission once an outage occurs. The maximum allowable latency for the periodic meter reading service is set to 300 seconds, the minimum reliability requirement is set to 99%, and the estimated data volume is approximately 32KB of the summary message of all node data within each meter reading cycle. This service has a medium transmission priority and can be completed within the scheduling window. The maximum allowable latency for the power quality waveform upload service is set to 60 seconds, the minimum reliability requirement is set to 95%, and the estimated data volume is approximately 8KB of waveform data per sample. This service has certain bandwidth requirements but allows for a certain degree of transmission failure and retransmission. The above transmission requirement vectors are stored locally on the acquisition terminal in the form of a configuration table, and can be updated through the main station when a new service type is added.

[0052] At the start of each scheduling cycle, the acquisition terminal first obtains the quality level sequence of each channel within the future prediction window from step S1, and simultaneously scans the current queue of services to be transmitted to obtain the type of each service and its corresponding transmission demand vector. Then, a two-stage decision-making mechanism is executed to generate a service-to-channel allocation scheme. The first stage is the constraint satisfaction stage, which traverses the queue of services to be transmitted from high to low priority, checking whether the predicted quality level of each available channel for each service meets the constraints in the service's transmission demand vector. Specifically, for the power outage event reporting service, the candidate channel's quality level within the prediction window must be no lower than "good" to meet its 99.9% reliability requirement and 3-second delay constraint; for the periodic meter reading service, the candidate channel's quality level must be no lower than "medium"; for the power quality waveform upload service, the candidate channel's quality level must be no lower than "good" and remain stable within the prediction window. After constraint checks, a set of candidate channels that meet the transmission demand constraints is selected for each service. If the candidate channel set for a certain service is empty, the service is marked as pending retry and re-evaluated in the next scheduling cycle.

[0053] The second stage is the global optimization stage. Based on the candidate channel set determined in the first stage, the optimization objective is to minimize the weighted transmission completion time of all services. A dynamic programming method is used to solve for the allocation scheme of each service to each channel. Specifically, services are prioritized and considered sequentially. For each service, the expected transmission completion time when allocated to each channel in the candidate set is evaluated. This time is calculated by dividing the estimated data volume by the available transmission rate of the channel at the corresponding quality level. The weight of each service is proportional to the reciprocal of its maximum permissible delay; that is, the weight of the power outage event reporting service is 1 / 3, the weight of the periodic meter reading service is 1 / 300, and the weight of the power quality waveform upload service is 1 / 60. The dynamic programming method searches for the scheme that minimizes the total weighted transmission completion time among all feasible allocation combinations. When the number of services and channels is limited (2 channels in this embodiment), the computational complexity of dynamic programming is controllable and can be solved in milliseconds.

[0054] Furthermore, the two-stage decision-making mechanism in this embodiment also introduces a channel handover cost evaluation mechanism. The acquisition terminal maintains a historical statistics table of inter-channel handover delays, recording the delay consumed by each handover operation between the HPLC channel and the wireless channel, and continuously updates the mean and standard deviation in the table. When the decision result of the global optimization stage requires a service to switch from the currently used channel to another channel, the average handover delay in that switching direction is read from the historical statistics table as the handover cost, and the reduction in the expected transmission completion time of the service after the handover compared to the case without handover is calculated as the handover benefit. A handover operation is only performed when the handover benefit is greater than 1.5 times the handover cost; this 1.5 times safety margin is used to cope with fluctuations in handover delays. This mechanism effectively avoids the problem of additional delay overhead introduced by frequent channel handovers, ensuring the stability of scheduling decisions.

[0055] like Figure 5As shown, using a typical 24-hour workday as the timeline, the diagram illustrates the causal time-series relationship between electricity consumption characteristics, channel quality prediction, and service scheduling. The vertical dashed line marks the critical moment of 7:30 AM, traversing all three panels to form a causal alignment. The upper panel displays the time-series curve of the total active power of the distribution area, showing a significant power surge around 7:30 AM when industrial loads start up, and a peak in residential electricity consumption around 6:00 PM. The middle panel shows the heatmap of quality level predictions for the HPLC and wireless channels, with quality levels categorized as excellent, good, medium, and poor. During the industrial load startup period, the HPLC channel experiences quality degradation due to power line noise interference, dropping from excellent to medium or poor. The wireless channel maintains relatively stable good quality throughout the day. The lower panel shows the Gantt chart for channel allocation for three types of services. Power outage reporting services are always prioritized for allocation to the wireless channel to ensure real-time performance. Periodic meter reading and waveform upload services use the HPLC channel during normal periods, migrating to the wireless channel in advance when HPLC channel quality is predicted to deteriorate. The three panels, aligned by vertical dotted lines, visually illustrate the causal relationship between load surges, channel degradation prediction, and preventative service migration. Compared to passive handover solutions, this invention initiates the prediction process immediately upon detecting load anomalies at 7:15, completing preventative service migration before 7:45 and channel handover approximately 15 minutes earlier. This avoids data loss caused by sudden channel quality drops during the 7:30-7:45 window, which is common in passive solutions.

[0056] like Figure 6 As shown, a two-dimensional decision space for service requirements and channel capabilities is constructed with transmission delay as the horizontal axis and transmission reliability as the vertical axis. The figure uses dashed rectangles to mark the constraint domains for the three types of services: the power outage reporting service is located in the compact area in the upper left corner, requiring a delay of less than 3 seconds and a reliability higher than 99.9%; the waveform uploading service is located in the central area, requiring a delay of less than 60 seconds and a reliability higher than 95%; the periodic meter reading service has the most lenient constraint domain, with a delay tolerance of up to 300 seconds. The scatter plots in the figure represent the measured performance of the HPLC and wireless channels at different quality levels. Solid dots represent the HPLC channel, and hollow dots represent the wireless channel. Different marker shapes correspond to four quality levels: excellent, good, medium, and poor. In the first stage of constraint satisfaction, the system selects candidate channels falling within each constraint domain according to service priority. In the second stage of global optimization, dynamic programming is used to solve for the allocation scheme with the minimum weighted transmission completion time. The figure also marks a handover safety margin line; a handover operation is only performed when the channel handover benefit exceeds the handover cost, avoiding frequent invalid handovers.

[0057] Step S3: When the number of power outage event reporting requests exceeds the storm threshold within the preset time window, the power outage events are classified into three modes based on the time diffusion rate and spatial diffusion range of the power outage events: local power outage, branch power outage, and whole-area power outage. The communication resources are then partitioned and reconstructed according to the power outage mode.

[0058] like Figure 3 As shown, in this embodiment, the storm threshold is set to 10% of the total number of nodes in the distribution area when the number of power outage event reporting requests received within 15 minutes reaches 10%. That is, when 26 or more power outage reporting requests are received within 15 minutes, the power outage event pattern recognition process is triggered. The acquisition terminal sorts all power outage reporting requests received within a preset time window according to the reporting time and extracts two features: the time diffusion rate and the spatial diffusion range of the power outage event. The time diffusion rate is calculated as follows: the time series of power outage reporting requests is divided into time slots with a 1-minute interval, and the number of newly reported power outage nodes in each time slot is counted. The maximum value is taken as the time diffusion rate (unit: nodes / minute). The spatial diffusion range is calculated as follows: all nodes that have reported power outages are marked in the distribution area topology map pre-stored in the acquisition terminal. The coverage range of the connected subgraph formed by these nodes is calculated using a depth-first search algorithm, and is measured by the number of covered branch lines and the proportion of covered nodes to the total number of nodes in the distribution area.

[0059] Based on two characteristics—temporal diffusion rate and spatial diffusion range—a rule engine is used to classify outage modes. In this embodiment, the first threshold is set to 3 nodes / minute, and the second threshold is set to 12 nodes / minute. When the spatial diffusion range is limited to a single branch line (i.e., the outage nodes are only distributed on the same branch line in the transformer area topology) and the temporal diffusion rate is less than 3 nodes / minute, it is determined to be a partial outage mode. When the spatial diffusion range covers two or more but not all branch lines, it is determined to be a branch outage mode. When the spatial diffusion range covers the entire transformer area (i.e., all branch lines have outage nodes) or the temporal diffusion rate exceeds 12 nodes / minute, it is determined to be a whole transformer area outage mode. The above rule determination checks the whole transformer area outage, branch outage, and partial outage in that order, and outputs the corresponding outage mode once a rule is met.

[0060] Based on the identified power outage mode, the corresponding communication resource partitioning and reconstruction strategy is executed. For localized power outage modes, since the HPLC channel in the outage area may be invalid, all reporting services from nodes within the outage area are rerouted to the wireless channel for transmission, while nodes in non-outage areas maintain the conventional scheduling scheme determined in step S2. This strategy minimizes disruption to the overall system scheduling, only making targeted adjustments to the affected areas.

[0061] For branch outage mode, the outage involves a large number of branch lines, increasing the access contention pressure on the wireless channel. In this case, the available capacity of the wireless channel is equally divided according to the number of outage branches, and an independent reporting time slot window is allocated to each outage branch. Specifically, if K branch lines are currently experiencing outages, each scheduling cycle of the wireless channel is equally divided into K time slot windows, with the duration of each time slot window being the scheduling cycle duration divided by K. Within each time slot window, outage reporting nodes within the corresponding branch are sorted in ascending order according to the CRC16 hash value of their node address (6-byte address in hexadecimal representation), and sequentially access the wireless channel to report the outage. This deterministic sorting mechanism avoids collisions and retransmissions caused by random node access contention, maximizing the utilization efficiency of the wireless channel.

[0062] For a full-area power outage, available communication resources are most strained, necessitating a tiered compression reporting strategy. First, all wireless channel capacity is allocated to power outage reporting, suspending low-priority services such as periodic meter readings and power quality waveform uploads. Second, the power outage reporting messages are simplified. The original message contains multiple fields including node identifier, outage timestamp, last voltage sample value, pre-outage load data, and device status word. After simplification, only three essential fields are retained: node identifier (6 bytes), outage timestamp (4 bytes), and last voltage sample value (2 bytes). A single simplified message is 12 bytes long, plus frame header and trailer overhead, totaling 20 bytes. Furthermore, a topology-based aggregation reporting mechanism is implemented between nodes. In the topology map, upstream nodes with remaining power supply (i.e., nodes not yet out of power) act as proxy nodes, collecting outage information from downstream out-of-power nodes through their still-available HPLC channels. The proxy node packages the power outage information collected from multiple downstream nodes into an aggregated message, which is then reported to the data acquisition terminal as a single message via a wireless channel. A single aggregated message can contain simplified power outage information from up to 50 nodes, with a total message length not exceeding 1024 bytes. Through this aggregation reporting mechanism, the number of access contention events on the wireless channel is reduced from the node level to the proxy node level, ensuring complete reporting of power outage range information with minimal available channel resources.

[0063] Step S4: Collect the actual transmission results of each service and compare them with the expected results during scheduling decisions. When the channel quality level prediction accuracy is lower than the preset lower limit, trigger the online incremental update of the channel quality prediction model.

[0064] In this embodiment, after each scheduling cycle, the acquisition terminal automatically records the actual transmission completion time, the actual channel identifier used, and the transmission success rate of each service within that cycle. The actual results are then compared item by item with the expected results calculated based on channel quality prediction during scheduling decisions. The calculation of the prediction deviation index includes two aspects: the channel quality level prediction accuracy is defined as the percentage of time periods where the actual quality level matches the predicted quality level; the service transmission completion time prediction error rate is defined as the absolute value of the difference between the actual transmission completion time and the expected transmission completion time divided by the expected transmission completion time. The acquisition terminal maintains a sliding window (the window length is the most recent 48 scheduling cycles, i.e., 12 hours) and continuously calculates the rolling average of the above deviation indexes within this window.

[0065] When the rolling mean of the channel quality level prediction accuracy falls below a preset accuracy lower limit (set to 80% in this embodiment), the online incremental update process of the channel quality prediction model is triggered. The incremental update employs the Elastic Weight Consolidation (EWC) method, which evaluates the importance of each model parameter to the learned knowledge using the Fisher information matrix. Specifically, before the incremental update, a diagonal approximation of the Fisher information matrix is ​​calculated based on the model parameters from the most recent complete training or the previous incremental update. The k-th diagonal element F_k in the Fisher information matrix represents the importance of the k-th model parameter theta_k to the learned task. During the incremental update, the loss function adds an EWC regularization term to the original weighted cross-entropy loss. The expression for the EWC regularization term is lambda / 2 * sum(F_k * (theta_k - theta_k_old)^2), where theta_k_old is the parameter value before the incremental update, and lambda is the regularization coefficient, set to 5000 in this embodiment. This regularization term constrains the update magnitude of important parameters. Parameters with larger F_k values ​​are subject to stronger constraints during updates, thus preventing catastrophic forgetting of historical knowledge during incremental updates. Incremental training samples are runtime data from the most recent 14 scheduling cycles (i.e., 3.5 hours). The learning rate for incremental training is set to one-fiftieth of the initial training learning rate, i.e., 4.6 × 10⁻⁻⁴. 6 The batch size for incremental training is set to 16, and the number of training epochs is set to 10. After incremental training is completed, the acquisition terminal replaces the original parameters with the updated model parameters, and the new model is immediately put into use for prediction in subsequent scheduling cycles.

[0066] Furthermore, the closed-loop feedback also includes adaptive adjustment of the power outage pattern recognition rule parameters in step S3. After each power outage event is processed, the acquisition terminal obtains confirmation information of the actual power outage range from the main station and compares the judgment result of pattern recognition in step S3 with the actual power outage range. If the pattern judgment result does not match the actual power outage range (e.g., the actual power outage is a full-area outage but it is judged as a branch outage), the storm threshold, the first threshold, and the second threshold are corrected according to a preset step size. Specifically, when a false negative occurs (the actual power outage range is greater than the judgment range), the corresponding threshold is reduced by 10%; when a false negative occurs (the actual power outage range is less than the judgment range), the corresponding threshold is increased by 5%. The asymmetric adjustment strategy, where the reduction is greater than the increase, is consistent with the asymmetric penalty design idea of ​​the loss function in step S1, both tending to conservatively estimate to reduce the risk of false negatives. After feedback correction of multiple power outage events, the storm threshold, the first threshold, and the second threshold will gradually converge to the optimal value that matches the actual operating characteristics of the transformer area, thereby continuously improving the classification accuracy of the power outage pattern.

[0067] Example 2

[0068] In a preferred embodiment of the present invention, a channel quality causal prediction mechanism based on the timing characteristics of electricity consumption behavior is described in detail. This embodiment focuses on the construction method of the channel quality prediction model, the model training strategy, and the coupling method between the model and scheduling decision.

[0069] The data acquisition terminal in the distribution area reads raw operation records from the local historical operation database using a sliding time window. The time window length is set to the past 7 to 30 days, with a step size of 15 minutes. After the data is read, it is cleaned to remove missing records caused by abnormal reasons such as terminal restarts or communication interruptions. For missing time periods, the average of adjacent time periods of the same type of day is used for interpolation to ensure the continuity of time series data.

[0070] Specifically, the components of the electricity consumption behavior time-series feature vector are calculated as follows: The total active power of the distribution area is divided into 96 15-minute time periods. The mean, peak value, and standard deviation of the instantaneous total active power collected in each time period are calculated. These three statistics are concatenated to form a one-dimensional vector of length 288 as this component. The load equipment start-stop event count sequence uses the cumulative number of switching actions of each type of load equipment in each time period as its element. Equipment types include three main categories: industrial load, commercial load, and residential load. The corresponding count sequences are concatenated to form a vector of length 288. The power factor change rate sequence calculates the difference in power factor between adjacent time periods, reflecting the rate of reactive load fluctuation. It has a length of 95. The daily type identifier uses a one-hot encoding method, classifying the date type into three categories: weekday, weekend, and statutory holiday, forming an encoding vector of length 3. The concatenation of the above components forms the final electricity consumption behavior time-series feature vector with a total dimension of 674.

[0071] Furthermore, the historical channel quality tags are obtained as follows. Historical quality indicators for the HPLC channel include the average signal-to-noise ratio (SNR) and frame error rate for each 15-minute period. Historical quality indicators for the wireless channel include Received Signal Strength Indicator (RSSI) and Round-Trip Time (RTT). An excellent HPLC channel quality level is defined as an SNR of at least 20 dB and a frame error rate of no more than 1%. A good level is defined as an SNR between 10 dB and 20 dB and a frame error rate between 1% and 5%. A medium level is defined as an SNR between 5 dB and 10 dB or a frame error rate between 5% and 15%. A poor level is defined as an SNR below 5 dB or a frame error rate exceeding 15%. Wireless channels are defined as follows: Excellent (RSSI not lower than -75dBm and RTT not exceeding 100ms); Good (RSSI between -85dBm and -75dBm and RTT between 100ms and 300ms); Medium (RSSI between -95dBm and -85dBm or RTT between 300ms and 1000ms); and Poor (RSSI below -95dBm or RTT exceeding 1000ms). The time-series feature vectors of electricity consumption behavior are aligned with the aforementioned channel quality level labels using 15-minute timestamps to form supervised training sample pairs. The ratio of training set to validation set is 8:2.

[0072] The channel quality prediction model employs an encoder-decoder architecture. The encoder consists of two stacked layers of gated recurrent units (GRUs), each with a hidden state dimension of 128 to 256. This encoder progressively encodes the input sequence of electricity consumption behavior time-series feature vectors, extracting the temporal dependencies. Preferably, the encoder input sequence length is set to the electricity consumption behavior feature vectors spanning the past 4 to 12 time periods (corresponding to 1 to 3 hours). The decoder introduces a temporal attention layer at each prediction time step, which performs weighted aggregation of the hidden state vectors from all time steps of the encoder. Specifically, the attention weights of the temporal attention layer are calculated as follows: The hidden state vector h_d of the decoder at the current time step is concatenated with the hidden state vector h_i of the encoder at the i-th time step. After passing through a fully connected layer, the energy value score_i is obtained. A normalized exponential function (softmax) is applied to score_i for all time steps to obtain the normalized attention weight alpha_i. Finally, the context vector c_t is obtained by weighting and summing all encoder hidden states according to alpha_i. c_t is concatenated with h_d and then passed through a classification layer to output the channel quality level probability distribution for the current time step. The above formula for calculating the attention weights is expressed in plain text as: alpha_i = exp(score_i) / sum_j(exp(score_j)), where score_i = tanh(W_a * concat(h_d, h_i) + b_a), W_a is the learnable attention parameter matrix, and b_a is the bias vector.

[0073] The model training employs a weighted cross-entropy loss function to address the asymmetric requirements of prediction direction. Let the true quality level be y, and the predicted quality level be y_hat. When the prediction direction is from good to deterioration (i.e., y_hat is better than y), the loss weight for the corresponding sample is set to w_pos; when the prediction direction is from deterioration to good (i.e., y_hat is worse than y), the loss weight for the corresponding sample is set to w_neg. The ratio of w_pos to w_neg is set to 2 to 5, preferably 3. This asymmetric weight design makes the model tend to conservatively estimate channel quality during prediction, imposing a higher penalty for missed channel degradation events, thereby reducing the probability of service transmission failures during the channel degradation window. The plain text expression for the weighted cross-entropy loss function is: L = -sum_n(w_n * sum_k(y_nk * log(p_nk))), where n is the sample index, k is the quality level index, y_nk is the one-hot encoded value of the true label, p_nk is the predicted probability output by the model, and w_n is the weight coefficient of the prediction direction corresponding to the sample.

[0074] The Adam optimizer was used during training, with an initial learning rate of 0.001 to 0.005, batch size of 32 to 128, and a total number of training epochs of 50 to 200. An early stopping mechanism based on validation set loss was introduced; training was stopped early if the validation set loss did not decrease within 10 consecutive epochs to prevent overfitting. The training hardware configuration consisted of a single graphics processing unit (GPU) with at least 8GB of VRAM. After training, the model's quality grade prediction accuracy was evaluated on the validation set, requiring an overall accuracy of at least 80% and a recall of at least 85% for poor grades.

[0075] Furthermore, the output of the channel quality prediction model is a sequence of channel quality levels within a preset future time window. The prediction time window length is set to 15 to 60 minutes, preferably 30 minutes. This prediction result is transmitted to the joint scheduling decision module at a time-step granularity, enabling the scheduling decision to proactively switch communication modes before the actual channel quality deteriorates. When the predicted quality level of a channel is expected to decrease from good or better to below medium within a future time step, the scheduling decision module initiates preparations for switching to a backup channel in advance. This includes initiating a link detection handshake on the backup channel and buffering pending messages in the current transmission queue, reducing the switching latency from the traditional passive response level of seconds to the level of hundreds of milliseconds.

[0076] When abnormal power consumption behavior exists within a distribution area (such as a sudden increase in harmonic distortion rate due to a large number of users simultaneously connecting to nonlinear loads), this power consumption behavior characteristic occurs infrequently in historical data, which may lead to increased prediction deviations in the prediction model for corresponding channel quality changes. Under the above boundary conditions, the impact of prediction uncertainty on transmission reliability can be compensated by increasing the channel quality safety margin in scheduling decisions (i.e., requiring the prediction quality level to be one level higher than the service demand level).

[0077] This implementation uses the temporal characteristics of power consumption behavior in the distribution area as the causal input for channel quality prediction, rather than relying solely on extrapolation from historical channel indicators. This fundamentally establishes a causal relationship between load behavior changes and channel degradation. On one hand, power consumption behavior characteristics have a certain leading effect on channel quality changes; sudden load increases are often accompanied by an earlier rise in HPLC channel noise. Therefore, predictions based on power consumption behavior can detect degradation trends several minutes earlier than the delayed response of channel indicators, providing a time margin for proactive handover. On the other hand, the introduction of daily type identifiers and time-period statistical features enables the model to capture periodic power consumption patterns. It exhibits high predictive stability for regular channel quality changes occurring at fixed times each day (such as concentrated residential load access during rush hour), with a lower false alarm rate than prediction methods that purely rely on historical channel indicators. The asymmetric design of the weighted cross-entropy loss function effectively suppresses the model's false negative rate (misclassifying degraded channels as good), reducing the service transmission failure rate while maintaining similar prediction accuracy.

[0078] The quality changes in the HPLC channel of low-voltage power distribution areas are physically closely related to the load status within the area. When high-power motor loads start up in the area, the amplitude of high-frequency noise on the power lines increases accordingly, and the signal-to-noise ratio of the HPLC channel decreases in a short period of time. Such load start-up and shutdown events are clearly reflected in both the power factor change rate sequence and the load start-up and shutdown count sequence, and their occurrence patterns exhibit statistical regularity based on intraday time periods and date types. The gated recurrent unit can selectively retain historical information useful for long-term trends and short-term fluctuations through forget gates and update gates, making it suitable for encoding such electricity consumption behavior characteristics with multi-scale time-series dependencies. The time attention mechanism dynamically calculates the contribution weight of each historical time step to the current prediction, enabling the model to adaptively focus on historical time periods most similar to the electricity consumption pattern at the current prediction time, avoiding the dilution of effective information caused by the uniform decay of long-sequence historical information. The asymmetric weighted loss function explicitly encodes the service transmission reliability requirements into the model training process at the optimization objective level, so that the model's risk preference is consistent with the scheduling requirements in actual applications, which would rather conservatively predict good channels as deteriorated than miss deterioration events. This ensures the alignment of the model's predictive behavior with the service objectives from the training mechanism perspective.

[0079] Example 3

[0080] In another embodiment of the present invention, the topological location information of the transformer area nodes is introduced into the channel quality prediction stage, and spatiotemporal joint modeling is used to replace the prediction method based solely on time-series features in Embodiments 1 and 2, in order to address the impact of transformer area topology differences on channel quality distribution.

[0081] In Examples 1 and 2, the channel quality prediction model uses a uniform time-series characteristic input of electricity consumption behavior for all nodes within the transformer area, without distinguishing the differentiated channel responses of nodes at different topological locations under the same load change. In actual transformer areas, the signal-to-noise ratio (SNR) of the HPLC channel at end nodes that are far from the transformer's electrical distance is usually lower than that of nodes on the main line. Furthermore, due to the influence of line impedance, the same load disturbance event has different degrees of impact on the channel quality of nodes at different locations. This implementation introduces node topological location features during the feature extraction stage and incorporates a graph attention network into the model structure to spatially encode the transformer area topological relationships, thereby achieving joint modeling of the temporal and spatial distribution patterns of channel quality.

[0082] Specifically, the topological location features of the transformer substation are calculated as follows: The substation topology is retrieved from the substation archive database. A topology map is constructed with the transformer as the root node and each acquisition node as a leaf node. Each edge in the map corresponds to a branch line, and the weight of the edge is set to the per-unit impedance value of that line segment. For each acquisition node in the substation topology map, the following topological location features are extracted: the shortest path electrical distance from the node to the transformer root node (measured by the sum of impedance path lengths), the number of branch nodes traversed by the node, and the degree of the node in the topology map (i.e., the number of other nodes directly connected to the node). These three scalar features are concatenated to form the topological location feature vector of the node, with a dimension of 3.

[0083] Furthermore, the model's input is constructed as follows: For each data acquisition node in the transformer substation, its electricity consumption behavior time-series feature vector is concatenated with its corresponding topological location feature vector to form an extended feature vector, which serves as the initial node feature for that node. Using the transformer substation topology map as the graph structure input, the extended feature vectors of each node are used as the initial node embeddings for the graph neural network.

[0084] In terms of model structure, this implementation adopts a cascaded structure of a graph attention network and a long short-term memory network, replacing the GRU encoder-decoder structure in Example 2. The spatial encoding stage consists of two graph attention network layers. Each layer aggregates the features of its neighboring nodes for each target node. The aggregation weights are dynamically calculated by the feature similarity of the node pairs through an attention mechanism to distinguish the degree of influence of different neighboring nodes on the channel quality of the target node. The formula for calculating the attention coefficient of the graph attention network layer is expressed in plain text as: alpha_ij = softmax_j(LeakyReLU(a^T * [W*h_i || W*h_j])), where h_i and h_j are the feature vectors of node i and node j, respectively, W is a learnable linear transformation matrix, a is the attention parameter vector, || denotes vector concatenation, and LeakyReLU is a linear rectified function with a negative slope, preferably 0.2. The output of the two graph attention network layers is a node embedding vector containing spatial topology information. Preferably, the output dimension of each graph attention network layer is set to 64 to 128.

[0085] The temporal coding stage receives the node embedding vector sequence (arranged by time step) output from the spatial coding stage, and a two-layer Long Short-Term Memory (LSTM) network performs temporal modeling on this sequence. The LSTM unit, through the coordinated action of the input gate, forget gate, and output gate, retains historical states valid for prediction over a relatively long time span; the preferred LSTM hidden state dimension is set to 128 to 256. The output of the temporal coding is then processed through a fully connected layer and activated by softmax to obtain the probability distribution of each channel quality level within the prediction time window.

[0086] The training dataset in this embodiment, based on the dataset described in Example 2, additionally includes the topology map structure information of the transformer substations and the topological location feature annotations of each node. The training data must cover the running data of at least three transformer substations of different sizes (approximately 50 nodes for small substations, approximately 100 to 300 nodes for medium-sized substations, and approximately 300 to 500 nodes for large substations) to enable the model to learn the generalization ability of topological location features across substation scales. During training, batches are constructed on a node-by-node basis, with a batch size of 32 to 64 nodes. The optimizer used is AdamW, with a weight decay coefficient of 0.01 to 0.05, an initial learning rate of 0.0005 to 0.002, and a cosine annealing scheduling strategy. The number of training epochs is 100 to 300.

[0087] Furthermore, during the scheduling decision-making phase, this implementation identifies the set of neighboring nodes with the greatest impact on the channel quality of the target node based on the spatial attention weights output by the graph attention network. The real-time power consumption changes of nodes within this set are used as real-time event sources to trigger channel quality re-prediction. When the real-time total active power of any node within the set deviates from the historical average for the current time period by more than 20%, the scheduling module immediately triggers an incremental prediction, updating the prediction sequence with the latest power consumption observations, correcting the original channel quality level prediction results, and adjusting the service allocation scheme to the channel accordingly. This mechanism enables the scheduling system to respond promptly to sudden load change events, compensating for the lag of fixed prediction cycles in sudden scenarios.

[0088] In power outage emergency scenarios, this implementation utilizes a transformer substation topology map to track the power outage spread path in real time. When a power outage report is detected, a breadth-first search is performed on the transformer substation topology map, starting from the already out-of-power node, to calculate the coverage area and spread boundary of the connected subgraph of the current out-of-power node set. This replaces the static rule judgment that relies on pre-stored paths in the topology map in Embodiment 1, allowing the power outage pattern recognition results to be dynamically updated as the power outage spreads in real time. After the power outage pattern recognition is completed, the communication resource partitioning and reconstruction strategy remains consistent with Embodiment 1, still implementing differentiated resource allocation schemes according to three modes: partial power outage, branch power outage, and full transformer substation power outage.

[0089] In the closed-loop feedback update phase, this implementation method, based on the elastic weight consolidation incremental update method described in Example 1, incorporates the topological position weights of the graph attention network into the calculation scope of the Fisher information matrix, applying the same constraints to the updates of topology-related parameters as to the temporal parameters. When a change in the topology of a transformer substation occurs (such as the addition of a new acquisition node or line modification), a complete retraining of the graph attention network layer parameters is triggered after the local substation topology graph database is updated. The LSTM layer parameters retain the learned temporal patterns through incremental fine-tuning, avoiding the overall model retraining overhead caused by topology updates.

[0090] When the transformer area topology map data is incomplete or contains incorrect annotations (such as missing line impedance parameters), this implementation method degenerates into the pure time-series prediction mode of Example 1, replacing the missing topology information with a zero vector of topology location features. At this time, the spatial aggregation of the graph attention network layer degenerates into uniform aggregation, which is consistent with the prediction using only time-series features in terms of effect, ensuring that the system can still operate normally when the data quality is insufficient.

[0091] Example 4

[0092] like Figure 4 As shown, this embodiment provides a multi-mode transmission and scheduling system for a low-voltage distribution transformer intelligent acquisition terminal. This system is deployed on a multi-mode communication acquisition terminal on the low-voltage distribution transformer side and runs as a software functional module within the terminal's embedded processor. The processor is an ARM architecture processor supporting floating-point operations, with a main frequency of no less than 800MHz, equipped with no less than 256MB of RAM to support inference calculations for the channel quality prediction model, and no less than 2GB of local storage space for storing historical operating data and model parameter files. The system includes a power consumption behavior feature extraction module, a channel quality prediction module, a service demand modeling module, a joint scheduling decision module, a power outage emergency reconstruction module, and a closed-loop feedback update module. The modules exchange data through an inter-process communication mechanism, forming an information flow path centered on scheduling decisions.

[0093] The electricity consumption behavior feature extraction module extracts time-series feature vectors of electricity consumption behavior from historical operating data collected by the distribution area acquisition terminals. These feature vectors include time-segmented statistical values ​​of the total active power of the distribution area, a sequence of start-stop event counts for load equipment, a power factor change rate sequence, and a daily type identifier. Specifically, this module reads raw operating data such as three-phase voltage, three-phase current, active power, reactive power, and power factor of the low-voltage side of the distribution area transformer from the terminal's local historical database according to the scheduling cycle. Missing values ​​caused by communication interruptions or sensor failures are filled using linear interpolation. Outliers exceeding the physically reasonable range are replaced with the mean of adjacent valid values. After cleaning, each dimension of the feature is standardized with zero mean and unit variance to eliminate the impact of dimensional differences between different physical quantities on subsequent model inference. The extracted electricity consumption behavior time-series feature vector is then transmitted to the channel quality prediction module via an inter-process communication interface.

[0094] The channel quality prediction module aligns the time-series feature vector of electricity consumption behavior with the historical quality indicators of various communication channels during the same period. Then, it outputs the quality level sequence of each communication channel within a preset future time window using a channel quality prediction model based on a recurrent neural network with a time attention mechanism. Specifically, the module loads the channel quality prediction model parameters stored locally on the terminal in file format. The model adopts an encoder-decoder architecture, with both the encoder and decoder employing two layers of gated recurrent units (GRUs). A time attention layer is introduced at each prediction time step of the decoder, and the hidden state sequences of each time step of the encoder are weighted and aggregated using additive attention calculation. This allows the model to adaptively focus on the historical time-series segments that have the greatest impact on the current prediction window, extracting periodic patterns and trend features of electricity consumption behavior. After receiving the feature vector from the electricity consumption behavior feature extraction module, the module calls the model inference interface to complete a forward calculation, outputting the predicted quality level sequences for the HPLC channel and the wireless channel within the preset future time window. The quality levels are divided into four discrete levels: excellent, good, medium, and poor. The prediction results are simultaneously pushed to the joint scheduling decision module and the power outage emergency reconstruction module in the form of structured messages for their use.

[0095] The service requirement modeling module defines transmission requirement vectors for each service type. These vectors include maximum permissible latency, minimum reliability requirements, and estimated data volume. This module maintains transmission requirement constraints for each service type locally on the terminal in the form of a configuration table, supporting remote updates of service requirement parameters by the master station via downlink configuration messages. At the start of each scheduling cycle, this module scans the current queue of services to be transmitted, matches the corresponding transmission requirement vector for each service in the queue, and transmits the type identifier, priority, and transmission requirement constraints of each service in the queue to the joint scheduling decision module in list form.

[0096] The joint scheduling decision module generates a service-to-channel allocation scheme based on the quality level sequence output by the channel quality prediction module and the transmission demand vector, using a two-stage decision mechanism. Specifically, the first stage is the constraint satisfaction stage. This module traverses the queue of services to be transmitted from high to low priority, checking whether the channel quality levels output by the channel quality prediction module for each service meet the delay and reliability constraints in the service's transmission demand vector, and selecting a set of candidate channels that meet the constraints. If the candidate channel set for a certain service is empty, the service is marked as pending retry. The second stage is the global optimization stage. Based on the candidate channel set determined in the first stage, and with the objective of minimizing the weighted transmission completion time of all services, a dynamic programming method is used to solve for the allocation scheme of each service to each channel. The weight of each service is proportional to the reciprocal of its maximum allowable delay. This module is also configured to perform channel handover cost assessment. It evaluates the handover cost using a maintained historical statistics table of handover delays between channels, and uses the reduction in the expected transmission completion time of the service after the handover as the handover benefit. The handover operation is only performed when the handover benefit exceeds a preset safety margin multiple of the handover cost, thus avoiding additional delay overhead caused by frequent channel handovers. The final determined service-to-channel allocation scheme is then sent and executed via the communication interface.

[0097] The power outage emergency reconfiguration module is used to classify power outage events into three modes—local power outage, branch power outage, and full-area power outage—based on the time diffusion rate and spatial diffusion range of the power outage events when the number of power outage event reporting requests exceeds the storm threshold within a preset time window. It then performs partitioned reconfiguration of communication resources according to the power outage mode. Specifically, this module continuously monitors the number of power outage event reporting requests arriving at the acquisition terminal. When it detects that the cumulative number of reported requests exceeds the storm threshold within the preset time window, the module takes over scheduling control and initiates the power outage mode identification process. During the identification process, the module calculates the spatial diffusion range based on the coverage area of ​​the connected subgraph of the already reported power outage nodes in the area topology map and counts the number of newly reported power outage nodes per unit time as the time diffusion rate. The rule engine sequentially checks the criteria for power outages in the entire distribution area, branches, and local areas. After outputting the power outage mode classification results, it executes the corresponding communication resource partitioning and reconstruction strategy: In the local power outage mode, all reporting services from nodes in the outage area are scheduled to the wireless channel; in the branch power outage mode, the wireless channel capacity is equally divided according to the number of outage branches, and an independent reporting time slot window is allocated to each branch; in the entire distribution area power outage mode, all wireless channel capacity is allocated to power outage reporting services, and the power outage reporting messages are simplified and encoded, retaining only three fields: node identifier, power outage timestamp, and last voltage sample value. Nodes with remaining power supply then report the power outage information of their downstream outage nodes in the form of aggregated messages. After the power outage event is processed, this module returns scheduling control to the joint scheduling decision module.

[0098] The closed-loop feedback update module collects the actual transmission results of each service and compares them with the expected results during scheduling decisions. When the channel quality level prediction accuracy is lower than a preset lower limit, it triggers an online incremental update of the channel quality prediction model. Specifically, at the end of each scheduling cycle, this module automatically collects the actual transmission completion time, actual channel usage, and transmission success rate of each service within that cycle, compares them item by item with the expected results during scheduling decisions, and continuously calculates the rolling average of the channel quality level prediction accuracy within a sliding window. When this rolling average is lower than a preset accuracy lower limit, it triggers an online incremental update process: using the Elastic Weight Consolidation (EWC) method, it evaluates the importance of each model parameter to the learned knowledge based on the Fisher information matrix, applies regularization constraints to the update magnitude of important parameters, and fine-tunes the channel quality prediction model parameters using the running data within the most recent preset number of cycles as incremental training samples to prevent catastrophic forgetting of historical knowledge during the incremental update process. After incremental training is completed, this module writes the updated model parameters to the terminal's local storage, which is then loaded and used by the channel quality prediction module in the next scheduling cycle. The module is also configured to adaptively correct the power outage pattern recognition rule parameters: after each power outage event is processed, the actual power outage range information fed back by the master station is compared with the judgment result of the pattern recognition in step S3. When a missed judgment occurs, the corresponding threshold is reduced by a preset ratio, and when a misjudgment occurs, the corresponding threshold is increased by a preset ratio, so that the storm threshold, the first threshold and the second threshold gradually converge to the optimal value that matches the actual operating characteristics of the transformer area.

[0099] Example 5

[0100] This embodiment illustrates the implementation process and technical effects of the multi-mode transmission scheduling method of the present invention in a real low-voltage distribution area operating environment through a specific application case.

[0101] The transformer substation described in this embodiment is located in a typical power distribution scenario where industrial plants and residential areas are supplied with mixed power. The substation has a total of 312 data acquisition nodes, including 128 industrial load nodes and 184 residential load nodes. The concentrated start-up and shutdown of industrial loads creates regular periods of strong interference on the substation's HPLC channel. The traditional passive handover scheme based on real-time channel quality detection has a measured channel availability rate of only 71.3% in this substation. During the morning and evening peak hours of concentrated industrial load start-up and shutdown, the frame error rate of the HPLC channel frequently exceeds 15%, resulting in a data loss rate of 8.7% for periodic meter reading services and a 12.4% rate of power outage event reporting delays exceeding 3 seconds. Existing general scheduling schemes are insufficient to meet the reliability requirements of this substation.

[0102] After adopting the method of this invention, step S1 is first executed to construct a training dataset from 120 days of historical operation data accumulated by the data acquisition terminals in the distribution area. The time-segmented characteristics of the total active power of the distribution area are statistically analyzed according to weekdays, weekends, and statutory holidays. It is found that industrial plants exhibit concentrated load start-up and shutdown characteristics during two time periods on weekdays: 7:30-8:30 and 17:00-18:30. The first-order differential peak value of active power during these time periods is 4.7 times the average value during non-production periods. Using the time-series feature vector of electricity consumption behavior containing the above characteristics as input, and the corresponding time period's HPLC channel signal-to-noise ratio and frame error rate, and wireless channel RSSI and RTT as labels, a total of 14,580 sets of input-label paired training samples are constructed. After dividing the training, validation, and test sets into a 7:1.5:1.5 ratio, the channel quality prediction model was trained on a server equipped with a single GPU using the PyTorch framework. The training process consisted of 87 rounds (with an early termination mechanism). The final model achieved a weighted F1 score of 0.879 on the test set, with a recall rate of 0.921 for channel quality degradation events, validating the effectiveness of the asymmetric penalty loss function design in improving the sensitivity of degradation event detection.

[0103] During actual scheduling and operation, the channel quality prediction model outputs a prediction that the HPLC channel quality will deteriorate to a poor level between 7:30 and 8:15 during the 7:15 scheduling cycle on weekdays (i.e., the cycle before the concentrated start-up of industrial loads), with a prediction lead of 15 minutes. Based on this, the joint scheduling decision module completes the preventative migration of the day's meter reading service to the wireless channel before 7:30, achieving uninterrupted continuous service transmission. Compared with the passive handover scheme based on real-time detection, the present invention eliminates data loss caused by the channel degradation window period (approximately 2 to 4 minutes) during this period, increases the integrity rate of periodic meter reading data from 91.3% to 98.6%, and reduces the proportion of power outage event reporting delays exceeding 3 seconds from 12.4% to 1.8%.

[0104] Furthermore, at 14:32 on a certain workday, a large-scale power outage occurred in the distribution area due to an upstream feeder fault. Within the first minute of the outage, the data acquisition terminal received outage reports from 47 nodes, exceeding the storm threshold (312 × 10% ≈ 32). The outage emergency reconstruction module then initiated outage mode identification. By calculating the connected subgraphs of each reported outage node on the distribution area topology, it was identified that the outage covered all four branch lines. Simultaneously, the number of newly reported nodes in the first minute was 47, far exceeding the second threshold (12 nodes / minute), thus classifying it as a full-area power outage. The power outage emergency reconfiguration module then implemented a resource reconfiguration strategy for the entire power outage mode: periodic meter reading and waveform upload services were suspended, and all wireless channel capacity was allocated to power outage reporting; power outage reporting messages were simplified and encoded, with each message compressed from approximately 96 bytes to 20 bytes; 12 upstream nodes with remaining power supply were identified as proxy nodes, which collected information from downstream outage nodes via the HPLC channel and reported it in aggregated message form. Within 15 minutes of the power outage, the system completed the reporting of power outage information for 305 outage nodes, achieving a power outage range information reporting completeness rate of 97.8%. Without the aggregated reporting mechanism (i.e., a scheme where nodes directly compete for access via wireless channels), the power outage information reporting completeness rate within 15 minutes was only 62.3% due to wireless channel access competition collisions, demonstrating the effectiveness of the aggregated reporting mechanism in the entire power outage scenario.

[0105] To verify the contribution of each component to the overall performance, an ablation experiment was also conducted in this embodiment. The ablation experiment was evaluated on data from 30 randomly selected working days within the same quarter in the aforementioned distribution area. Evaluation metrics included the predicted recall rate of channel quality degradation events, the completeness rate of periodic meter reading data, and the rate of power outage event reporting delay exceeding the standard. The experiment was grouped as follows: Scheme A was the complete method of this invention; Scheme B was an ablation scheme that removed the time attention mechanism (replacing the attention layer in the decoder with a non-attentional ordinary GRU decoder); Scheme C was an ablation scheme that removed the asymmetric penalty loss function (setting the asymmetric weights in the weighted cross-entropy to be equal); Scheme D was an ablation scheme that removed the online incremental update mechanism (model parameters are fixed and not updated); and Scheme E was a traditional passive handover scheme based on real-time channel quality detection (as a baseline comparison). The evaluation results are shown in the table below.

[0106] Option A (Complete Option): Channel quality degradation event prediction recall rate 92.1%, periodic meter reading data integrity rate 98.6%, power outage reporting latency exceedance rate 1.8%. Option B (Removing Time Attention): Channel quality degradation event prediction recall rate 84.3%, periodic meter reading data integrity rate 95.1%, power outage reporting latency exceedance rate 4.2%. Option C (Removing Asymmetric Penalty): Channel quality degradation event prediction recall rate 80.7%, periodic meter reading data integrity rate 94.3%, power outage reporting latency exceedance rate 5.1%. Option D (Removing Online Incremental Updates): Performance is similar to Option A for the first 15 days, but from day 16 to day 30, due to seasonal changes in the operating characteristics of the distribution area, the channel quality degradation event prediction recall rate drops to 80.2%, the periodic meter reading data integrity rate is 96.0%, and the power outage reporting latency exceedance rate rises to 3.9%, indicating the necessity of the online incremental update mechanism for maintaining long-term scheduling performance. Option E (passive switching baseline): Channel quality degradation event prediction recall rate 71.3% (equivalent to channel real-time availability rate), periodic meter reading data integrity rate 91.3%, and power outage reporting delay exceedance rate 12.4%.

[0107] like Figure 8 As shown in the figure, the horizontal axis represents five ablation schemes, the left vertical axis corresponds to the channel quality degradation prediction recall rate and periodic meter reading data integrity rate using a two-bar grouping method, and the right vertical axis corresponds to the power outage reporting latency exceedance rate using a diamond-marked line graph. Scheme A is the complete scheme, Scheme B removes the time attention mechanism, Scheme C removes the asymmetric penalty loss, Scheme D removes online incremental updates (16 to 30 days), and Scheme E is the passive switching baseline. Scheme A achieves the best results in all three metrics: recall rate 92.1%, meter reading integrity rate 98.6%, and latency exceedance rate 1.8%. After removing the time attention mechanism, the recall rate drops to 84.3%, indicating that this mechanism plays a key role in capturing the temporal patterns of electricity consumption behavior. Removing the asymmetric penalty loss further reduces the recall rate to 80.7%, indicating that the weighted cross-entropy effectively improves the sensitivity to asymmetric penalties for degradation events. Removing online incremental updates resulted in performance close to the complete solution for the first 15 days, but the recall rate dropped to 80.2% from days 16 to 30, validating the parameter maintenance role of the EWC mechanism in long-term operation. The passive switching baseline had the worst performance across all indicators, with a power outage reporting latency exceeding the standard rate of 12.4%, approximately 10.6 percentage points higher than Solution A.

[0108] like Figure 9As shown, the horizontal axis represents the time after a power outage (0 to 15 minutes), and the vertical axis represents the completeness rate of power outage information reporting. Three curves are plotted in the figure: solid circular markers represent the aggregation reporting scheme of this invention, dashed square markers represent the non-aggregation per-node competition scheme, and dotted triangle markers represent the theoretical upper limit. The shaded area indicates the performance difference between the two schemes. The scheme of this invention responds quickly after a power outage, achieving a completeness rate of approximately 60% at 5 minutes, approximately 85% at 10 minutes, and 97.8% at 15 minutes, approaching the theoretical upper limit. The non-aggregation per-node competition scheme is affected by channel collisions, resulting in a significantly slower rate of increase, reaching only 62.3% at 15 minutes. The curve flattens out in the later stages, exhibiting obvious collision saturation characteristics. The difference in completeness rate between the two schemes at 15 minutes is 35.5 percentage points, demonstrating the performance advantage of the simplified coding and proxy node aggregation reporting mechanism of this invention in high-concurrency power outage scenarios.

[0109] like Figure 10 As shown, the figure contains two heatmaps side-by-side. The horizontal axis represents the number of days of operation (day 1 to day 30), and the vertical axis represents six time periods within a day (0:00 to 4:00, 4:00 to 8:00, 8:00 to 12:00, 12:00 to 16:00, 16:00 to 20:00, and 20:00 to 24:00). The color scale transitions from green (high accuracy, close to 100%) to yellow and then to red (low accuracy, approximately 70%). The left subplot shows the results with EWC incremental updates. The overall color tone remains uniformly green throughout the 30-day operation, and the accuracy in each time period is stable above 85%. The overall average on day 30 is approximately 91%, indicating that the EWC mechanism effectively prevents performance degradation caused by load pattern drift. The right subplot shows the results without incremental updates. The color tone for the first 15 days is similar to the left subplot. From day 16 onwards, large areas of yellow and red gradually appear, with the most significant degradation during the morning and evening peak hours. The overall average on day 30 drops to approximately 80%, about 11 percentage points lower than the EWC solution. The above comparison quantitatively demonstrates the necessity and effectiveness of the online incremental update mechanism of this invention in maintaining prediction accuracy in long-term deployment scenarios.

[0110] The ablation comparison results above show that the time attention mechanism improves the recall rate of channel quality degradation event prediction by approximately 7.8 percentage points, the asymmetric penalty loss function further improves the recall rate by approximately 3.6 percentage points, and the online incremental update mechanism keeps the recall rate stable during long-term operation. The combined effect of these three techniques reduces the outage reporting delay exceedance rate by 10.6 percentage points compared to the passive switching baseline, effectively solving the technical problems of high channel quality prediction difficulty and low scheduling reliability in mixed industrial-residential load areas.

[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-mode transmission scheduling method for a low-voltage distribution area intelligent data acquisition terminal, characterized in that, Includes the following steps: Step S1: Extract the time-series feature vector of electricity consumption behavior from the historical operation data of the distribution area acquisition terminal. The time-series feature vector of electricity consumption behavior includes the time-segmented statistical value of the total active power of the distribution area, the start-stop event count sequence of the load equipment, the power factor change rate sequence, and the type identifier of the day. After aligning the time-series feature vector of electricity consumption behavior with the historical quality indicators of each communication channel during the same period, train a channel quality prediction model based on a recurrent neural network with a time attention mechanism, and output the quality level sequence of each communication channel within a future preset time window. Step S2: Define a transmission demand vector for each service type. The transmission demand vector includes the maximum allowable delay, the minimum reliability requirement, and the estimated data volume. Based on the quality level sequence and the transmission demand vector, a two-stage decision mechanism is used to generate a service-to-channel allocation scheme. Step S3: When the number of power outage event reporting requests exceeds the storm threshold within the preset time window, the power outage events are classified into three modes based on the time diffusion rate and spatial diffusion range of the power outage events: local power outage, branch power outage, and whole-station power outage. The communication resources are then partitioned and reconstructed according to the power outage mode. Step S4: Collect the actual transmission results of each service and compare them with the expected results during scheduling decisions. When the channel quality level prediction accuracy is lower than the preset lower limit, trigger the online incremental update of the channel quality prediction model.

2. The method according to claim 1, characterized in that, The encoder of the channel quality prediction model uses a gated cyclic unit to perform time-series encoding on the time-series feature vector of the electricity consumption behavior. The decoder introduces a time attention layer to perform weighted aggregation of the hidden states of the encoder at each time step. The time attention layer adopts an additive attention calculation method, and its attention weight is calculated by concatenating the hidden state of the encoder and the current hidden state of the decoder, mapping through a fully connected layer, and then calculating through a normalized exponential function.

3. The method according to claim 1, characterized in that, The rule for classifying power outage events into three modes based on the time diffusion rate and spatial diffusion range of power outage events is as follows: when the spatial diffusion range is limited to a single branch line and the time diffusion rate is lower than a first threshold, it is determined to be a local power outage; when the spatial diffusion range covers multiple branch lines but not all of them and the time diffusion rate does not exceed a second threshold, it is determined to be a branch power outage; when the spatial diffusion range covers the entire transformer area or the time diffusion rate exceeds the second threshold, it is determined to be a full transformer area power outage; wherein the time diffusion rate is defined as the maximum value of the number of newly reported power outage nodes in each unit time slot within a preset time window, and the spatial diffusion range is defined as the coverage range of the connected subgraph on the transformer area topology map of the reported power outage nodes.

4. The method according to claim 1, characterized in that, In the two-stage decision-making mechanism, the first stage is the constraint satisfaction stage, which traverses the queue of services to be transmitted from high to low according to service priority, and selects a set of candidate channels for each service whose predicted quality level satisfies the transmission demand vector constraint. The second stage is the global optimization stage, which aims to minimize the weighted transmission completion time of all services. The allocation scheme is solved by dynamic programming, and the weight of each service is proportional to the reciprocal of its maximum allowable delay.

5. The method according to claim 3, characterized in that, For the partial power outage mode, all reporting services from nodes within the outage area are scheduled to the wireless channel, while the non-outage area maintains the regular scheduling. For the branch power outage mode, the wireless channel capacity is equally divided according to the number of outage branches, and an independent reporting time slot window is allocated to each outage branch. The reporting order of outage reporting nodes within each branch is determined by the hash value of the node address. For the full-area power outage mode, all wireless channel capacity is allocated to the outage reporting service. The outage reporting message is simplified and encoded, retaining only three fields: node identifier, outage timestamp, and last voltage sample value. Nodes with remaining power supply act as proxies to collect outage information from their downstream outage nodes and report it uniformly in the form of aggregated messages.

6. The method according to claim 1, characterized in that, The channel quality prediction model employs a weighted cross-entropy loss function during training, assigning a higher penalty weight to prediction errors that occur when channel quality deteriorates than when it deteriorates.

7. The method according to claim 1, characterized in that, The online incremental update employs an elastic weight consolidation method, which applies regularization constraints to the learned important parameters when updating the model parameters, and uses the running data within the most recent preset number of cycles as incremental training samples to fine-tune the parameters of the channel quality prediction model.

8. The method according to claim 1, characterized in that, The two-stage decision-making mechanism also introduces a channel handover cost evaluation mechanism. When the decision result requires a certain service to be switched from the current channel to another channel, the handover cost is evaluated by maintaining the historical statistics table of handover delay between each channel. The reduction in the expected transmission completion time of the service after the handover is taken as the handover benefit. The handover operation is only performed when the handover benefit is greater than the preset safety margin multiple of the handover cost.

9. The method according to claim 1, characterized in that, The preset time window is 15 to 60 minutes, and the quality level in the quality level sequence is divided into four discrete levels: excellent, good, medium and poor. The historical quality indicators include the signal-to-noise ratio and frame error rate of the HPLC channel, as well as the received signal strength indicator and round-trip delay of the wireless channel.

10. A multi-mode transmission and scheduling system for a low-voltage distribution area intelligent data acquisition terminal, characterized in that, include: The electricity consumption behavior feature extraction module is used to extract the electricity consumption behavior time-series feature vector from the historical operation data of the distribution area collection terminal. The electricity consumption behavior time-series feature vector includes the time-segmented statistical value of the total active power of the distribution area, the start-stop event count sequence of the load equipment, the power factor change rate sequence, and the type identifier of the day. The channel quality prediction module is used to align the time-series feature vector of the electricity consumption behavior with the historical quality indicators of each communication channel during the same period, and then output the quality level sequence of each communication channel within a future preset time window through a channel quality prediction model based on a recurrent neural network with a time attention mechanism. The business requirement modeling module is used to define transmission requirement vectors for each business type. The transmission requirement vectors include the maximum allowable latency, the minimum reliability requirement, and the estimated data volume. The joint scheduling decision module is used to generate a service-to-channel allocation scheme based on the quality level sequence output by the channel quality prediction module and the transmission demand vector, using a two-stage decision mechanism. The power outage emergency reconstruction module is used to classify power outage events into three modes—local power outage, branch power outage, and whole-station power outage—based on the time spread rate and spatial spread range of the power outage events when the number of power outage event reporting requests exceeds the storm threshold within a preset time window, and to reconstruct communication resources in partitions according to the power outage mode. The closed-loop feedback update module is used to collect the actual transmission results of each service and compare them with the expected results during scheduling decisions. When the channel quality level prediction accuracy is lower than the preset lower limit, the online incremental update of the channel quality prediction model is triggered.

Images