Electric power communication network multi-link switching method and system based on link quality prediction

Abstract

The invention provides an electric power communication network multi-link switching method and system based on link quality prediction, and belongs to the technical field of electric power communication and Internet of Things access control. The method comprises the following steps: calculating a link instantaneous quality index based on link real-time characteristics; training a logistic regression model through a maximum likelihood estimation method to predict an offline risk quality index; fusing according to short-term disturbance characteristics and a cross-scale gating fusion method, inputting the short-term disturbance characteristics into an encoder-decoder to generate future link quality predicted values, and aggregating the future link quality predicted values into link trend quality indexes; fusing the link instantaneous quality index, the offline risk quality index and the link trend quality index into comprehensive link quality; and if the link switching condition is satisfied, determining the link with the highest comprehensive score according to the comprehensive link quality, switching to the link with the highest comprehensive score, and switching to the original link if the comprehensive link quality of the new link is continuously lower than a set threshold within a set time after switching. According to the invention, the stability and communication continuity of multilink access are improved.

Description

Technical Field

[0001] This invention relates to the field of power communication and Internet of Things access control technology, and more specifically, to a method and system for multi-link switching in power communication networks based on link quality prediction. Background Technology

[0002] With the advancement of the construction of the power Internet of Things (IoT), power terminal equipment, including but not limited to smart meters, feeder terminals, distributed monitoring devices, and low-power e-ink terminals in substations, are increasingly using wireless communication methods to access information interaction via ZigBee, LoRa, NB-IoT, 4G / 5G, or multi-mode wireless modules. However, different communication technologies exhibit significant differences in coverage, anti-interference capabilities, bandwidth, latency, and energy consumption. Furthermore, they are affected by environmental factors, including but not limited to metal reflections within the substation, electromagnetic interference from switching operations, and periodic vibrations from equipment operation. This results in link quality exhibiting distinct "multi-indicator, multi-scale, and multi-fluctuation characteristics." Traditional multi-link selection or switching methods mostly rely on a single communication indicator (such as RSSI or SNR) or a simple weighted scoring method. These solutions do not consider future link change trends, dropout probabilities, or latent interference patterns, and are typically unable to simultaneously adapt to the heterogeneous link characteristics of ZigBee / LoRa / 4G, leading to untimely switching, frequent false switching, or the inability to proactively mitigate link degradation risks. Existing technologies have proposed a method for predicting link quality using machine learning models, but they are mostly focused on single-link prediction, lack a unified evaluation mechanism to represent the quality of multiple communication links, and have not formed a "three-layer multi-scale link quality assessment system" that can be used in engineering field.

[0003] Prior art document 1 (CN103338472A) discloses a method for estimating the quality of a wireless network link. Its shortcomings lie in that this method primarily relies on statistical estimation of packet loss rate and latency parameters of a single wireless link based on the arrival time of probe data packets. It focuses on the parameterized description of the link's historical or current state, without addressing the unified quality characterization and comparative evaluation among multiple heterogeneous communication links. Furthermore, this method is a post-hoc estimation mechanism, unable to predict future link quality trends or potential downtime risks, and it does not integrate the link quality estimation results with the multi-link handover decision-making process. Therefore, it is difficult to directly apply to engineering handover scenarios that require proactive mitigation of link degradation risks.

[0004] Prior art document 2 (CN120897250A) discloses a communication link switching method and related apparatus. Its shortcomings lie in that this method mainly relies on multiple channel quality indicators within the current or preset time window to dynamically adjust the link switching threshold and switching judgment period. Its switching decision is essentially still based on reactive judgment of the current or historical channel quality status. Furthermore, this method does not model or predict future link quality trends and potential call drops. When link quality deteriorates rapidly in complex electromagnetic interference environments, switching lag or erroneous switching may still occur. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a multi-link handover method and system based on link quality prediction. By constructing a unified link quality index, estimating the maximum likelihood of connection loss risk, and improving the future trend prediction of Autoformer, it achieves early handover, jitter suppression, and failure fallback. This invention enables unified and comparable quality evaluation among heterogeneous links and performs "early avoidance handover" based on future trends, effectively improving communication continuity.

[0006] The present invention adopts the following technical solution.

[0007] The first aspect of the present invention provides a multi-link handover method for power communication networks based on link quality prediction, comprising: Based on the real-time characteristics of the link, the instantaneous quality index of the link is calculated by weighted normalization; Based on historical link features and link status labels, a logistic regression model is trained using the maximum likelihood estimation method. Real-time link features are then input into the trained logistic regression model to predict the quality index of disconnection risk. Short-term disturbance features are determined based on dynamic convolutional kernel weights based on the characteristics of the power communication network environment. Trend features are determined based on short-term disturbance features. Trend features and short-term disturbance features are fused using a cross-scale gating fusion method. The fusion result is input into an encoder-decoder to generate link quality prediction values. The link quality prediction values ​​are aggregated into a link trend quality index. The instantaneous link quality index, the disconnection risk quality index, and the link trend quality index are combined into a comprehensive link quality index. If the link switching conditions are met, the link with the highest comprehensive score is determined based on the overall link quality. If the link with the highest comprehensive score is not the link that was switched in the last time, the current link is switched to the link with the highest comprehensive score. If the link with the highest comprehensive score is the link that was switched in the last time, the current link is switched to the link with the highest comprehensive score when the score difference between the current link and the link with the highest comprehensive score is greater than a set threshold. If the overall link quality of the new link after switching is continuously lower than the difference between the overall link quality of the original link before switching and the fallback threshold within a set period of time, the original link is switched back.

[0008] Preferably, the instantaneous link quality index is expressed by the following formula:

[0009] In the formula, This represents the instantaneous quality index of the i-th link, where i=1 represents a ZigBee link, i=2 represents a LoRa link, and i=3 represents a 4G link. This represents the real-time characteristics of the i-th link at time t. The k-th normalized index, Indicates the number of normalization indicators. It represents the weight of the k-th normalized index.

[0010] Preferably, the predicted quality index for disconnection risk includes: Collect historical characteristics and link status labels of each link when it is in normal communication state and when it is disconnected to establish a logistic regression model, and train the logistic regression model by the maximum likelihood estimation method. The real-time collected link features are input into the trained logistic regression model to determine the probability that the link is available in the current state, which is set as the disconnection risk quality index.

[0011] Preferably, a logistic regression model is established, expressed by the following formula:

[0012] In the formula, Indicates link historical characteristics Probability of downlink availability, link history characteristics Includes normalized historical signal-to-noise ratio metrics Historical packet reception rate indicators Historical received signal strength indicators Historical round-trip delay indicators , Indicates the link status label, when =1 indicates that the link is in an available state. A value of 0 indicates that the link has dropped. This represents the Sigmoid function. For bias terms, , , and This represents the link feature weights.

[0013] Preferably, the calculation of the dynamic convolution kernel weights for the environmental characteristics of the power communication network includes: Based on the different environmental characteristics of power communication networks, an environmental probability density function of power communication networks is constructed using a Gaussian mixture model. The optimal parameters of the environmental probability density function are obtained by training the function using the expectation-maximization algorithm. Calculate the probability density value of each environmental category based on the optimal parameters, and sum the weighted probability density values ​​of each environmental category to determine the weighted probability density sum of all environmental categories. The posterior probability of the current environment belonging to each category is calculated by summing the weighted probability densities of all environment categories based on Bayes' theorem. The dynamic convolution kernel weights are calculated based on the posterior probability that the current environment belongs to each category.

[0014] Preferably, determining the characteristics of short-term disturbances includes: The link quality features within the historical time window are obtained. Within each sliding time window of the link quality features, short-term perturbation features are extracted using dynamic convolutional kernel weights, as expressed by the following formula:

[0015] In the formula, Indicates the dynamic convolution kernel weights. This indicates that time t is the endpoint and the length is... Link quality characteristics collected within the time window. This represents the perturbation bias term.

[0016] Preferably, generating link quality prediction values ​​includes: The short-term perturbation features are added to the original input as residuals to form the enhanced link quality features; The enhanced link quality features are decomposed into a sequence to obtain a trend vector; The trend vector is processed by an encoder to extract trend features. The trend features and short-term perturbation features are concatenated and the cross-scale fusion weights of the two features in the fusion result are dynamically adjusted by a learnable gating factor. Based on the cross-scale fusion weights, features of different scales are weighted and reconstructed to obtain cross-scale fusion features; The encoder uses an autocorrelation attention mechanism to encode cross-scale fused features, and the encoder output is passed through the decoder's multi-layer autocorrelation attention and cross attention to generate link quality prediction values.

[0017] Preferably, the cross-scale fusion feature is expressed by the following formula:

[0018] In the formula, This represents the cross-scale fusion feature at time t. Indicates trend characteristics, This represents the short-term perturbation features extracted at time t. The cross-scale fusion weights calculated at time t are represented by the following formula:

[0019] Learnable gating factors include learnable weight matrices. and learnable bias vector , This represents the Sigmoid function.

[0020] Preferably, the link with the highest overall score is determined based on the overall link quality, expressed by the following formula:

[0021] In the formula, This indicates the link with the highest overall score. The overall score of the i-th link is represented by the following formula:

[0022] In the formula, , and These represent the weighting coefficients for quality, delay, and cost, respectively. To improve overall link quality, This represents the latency metric for the i-th link. This represents the communication cost of the i-th link.

[0023] A second aspect of the present invention provides a multi-link handover system for power communication networks based on link quality prediction, which, when running the multi-link handover method for power communication networks based on link quality prediction described in the first aspect, includes: The link instantaneous quality index calculation module is used to calculate the link instantaneous quality index based on the link's real-time characteristics through weighted normalization. The disconnection risk quality index calculation module is used to train a logistic regression model based on historical link features and link status labels using the maximum likelihood estimation method, and input real-time link features into the trained logistic regression model to predict the disconnection risk quality index. The link trend quality index calculation module is used to determine short-term disturbance features based on the dynamic convolution kernel weights based on the characteristics of the power communication network environment, determine trend features based on the short-term disturbance features, fuse the trend features and short-term disturbance features through a cross-scale gating fusion method, input the fusion result into the encoder-decoder to generate link quality prediction values, and aggregate the link quality prediction values ​​into the link trend quality index. The fusion module is used to combine the instantaneous link quality index, the disconnection risk quality index, and the link trend quality index into a comprehensive link quality index. The switching module is used to determine the link with the highest comprehensive score based on the overall link quality if the switching conditions are met. If the link with the highest comprehensive score is not the link that was switched in the last time, the current link will be switched to the link with the highest comprehensive score. If the link with the highest comprehensive score is the link that was switched in the last time, the current link will be switched to the link with the highest comprehensive score when the score difference between the current link and the link with the highest comprehensive score is greater than a set threshold. If the overall link quality of the new link after switching is lower than the difference between the overall link quality of the original link before switching and the fallback threshold within a set period of time, the original link will be switched back.

[0024] Compared with the prior art, the beneficial effects of the present invention include at least the following: This application addresses heterogeneous multi-link access scenarios in power communication networks, including ZigBee, LoRa, NB-IoT, and 4G / 5G. It constructs a three-layer link quality assessment system consisting of instantaneous link quality assessment, dropout risk prediction, and future trend prediction. This system not only quantifies the current state of the link to achieve comparability between different communication technologies, but also uses a predictive model to pre-determine the future evolution trend and dropout probability of the link. The multi-layer assessment results are integrated to form a unified link quality index, which is used to directly guide multi-link selection and switching decisions, enabling comparability evaluation between links of different communication standards and avoiding switching decision bias caused by differences in the indicator system. This invention employs hierarchical modeling and fusion evaluation of instantaneous link quality, downtime risk, and future evolution trends. This enables link quality assessment to no longer rely solely on a single moment or indicator, but to simultaneously reflect the current state of the link, potential failure risks, and future trends, thereby achieving early detection and proactive avoidance of link degradation. Furthermore, by constructing an intelligent decision-making mechanism that includes handover triggering, candidate link selection, handover execution, jitter suppression, and failure fallback, the probability of erroneous and frequent handovers is effectively reduced, improving the stability and robustness of multi-link handover processes and significantly enhancing the continuity and reliability of power communication systems in complex field environments. This application introduces a drop risk prediction based on historical samples before link switching decisions to quantify the probability of a link dropping in the future. By predicting the drop risk and future trends, it enables early switching before link deterioration, thereby reducing the probability of communication interruption. Compared with the scheme that only performs parameter estimation, this application can significantly improve the stability and communication continuity of multi-link access in complex power field environments. By predicting future link quality trends through multi-scale sequence decomposition and improvements to the Autoformer encoder-decoder, a three-layer link quality assessment mechanism covering the current state, potential risks, and future trends is constructed. Based on the fused unified link quality index, prediction-driven intelligent multi-link switching is achieved. By predicting the future evolution trend of links, link switching is transformed from a reactive response to a proactive avoidance, effectively reducing the risk of communication interruption caused by sudden link deterioration. By fusing the results of multi-layer quality assessment, the stability and accuracy of switching decisions are improved in scenarios with frequent fluctuations in multiple links. This application can reduce unnecessary frequent switching, reduce system resource consumption, and improve the overall reliability of power communication networks. This invention is applicable to the access control of power Internet of Things (IoT) devices such as smart meters, power distribution terminals, and low-power monitoring devices between heterogeneous wireless links such as ZigBee, LoRa, NB-IoT, and 4G / 5G. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the multi-link switching method provided according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0027] like Figure 1 As shown, Embodiment 1 of the present invention provides a multi-link handover method for power communication networks based on link quality prediction, including the following steps: Step 1: Collect physical layer metrics of heterogeneous links in real time, map them to unified link metrics, normalize the link metrics to determine the real-time characteristics of the links; based on the real-time characteristics of the links, calculate the instantaneous link quality index through weighted normalization.

[0028] In a preferred but non-limiting embodiment of the present invention, step 1 includes: Step 1.1: Collect physical layer metrics of heterogeneous links in real time.

[0029] The heterogeneous links include ZigBee, LoRa, and 4G.

[0030] The physical layer metrics of the heterogeneous links include at least one of RSSI (Received Signal Strength Indicator), LQI (link quality indicator), SNR (Signal-to-Noise Ratio), PRR (Packet Reception Ratio), RTT (Round Trip Time), and SINR (Signal to Interference plus Noise Ratio).

[0031] Step 1.2: Map physical layer metrics to unified link metrics, and normalize the link metrics to determine unified real-time link characteristics.

[0032] More preferably, step 1.2 includes: ZigBee links collect RSSI and LQI, derive SNR based on LQI, and calculate PRR using empirical formulas based on LQI and RSSI values. ZigBee links determine RTT by actively sending probe packets and measuring round-trip time. The unified link metrics for ZigBee links include RSSI, SNR, PRR, and RTT.

[0033] LoRa links use directly collected RSSI, SNR, and PRR, and determine RTT by measuring round-trip time through the acknowledgment mechanism of the LoRaWAN protocol. The unified link metrics for LoRa links include RSSI, SNR, PRR, and RTT.

[0034] 4G link metrics include RSRP (Reference Signal Received Power), SINR, and packet loss rate. RSRP is equivalently mapped to RSSI, SNR is calculated based on SINR, and PRR is equivalently mapped to PRR. RTT is determined by sending Ping packets. The unified link metrics for 4G include RSSI, SNR, PRR, and RTT.

[0035] The range of each link metric is normalized. An example of normalization is shown below: (1) (2) Where RSSI represents the received signal strength indication value collected at the current moment; RSSImin and RSSImax represent the minimum and maximum received signal strength values ​​within a preset statistical time window, respectively; RTT represents the round-trip time delay measured at the current moment; RTT min and RTT max These represent the minimum and maximum round-trip delays within the statistical time window, respectively; RSSI norm and RTT norm These represent the normalized received signal strength index and round-trip delay index, respectively, with values ​​ranging from [0, 1], used to eliminate the influence of different dimensions and value ranges on subsequent link quality calculations.

[0036] The final unified real-time characteristics of the link are represented by the following formula: (3) In the formula, This represents the unified real-time characteristics of the i-th link at time t. In a heterogeneous network including ZigBee, LoRa, and 4G, i=1 represents a ZigBee link, i=2 represents a LoRa link, and i=3 represents a 4G link; SNR norm and PRR norm represents the signal-to-noise ratio and packet reception rate after normalization, respectively, and their values ​​range from [0, 1].

[0037] Unified real-time link features are used to uniformly characterize the instantaneous quality status of heterogeneous links at the same moment, and serve as input features for subsequent link quality assessment and intelligent switching decisions.

[0038] Step 1.3: Based on unified real-time link characteristics, calculate the instantaneous link quality index through weighted normalization, expressed by the following formula: (4) In the formula, This represents the instantaneous link quality index of the i-th link. This represents the real-time characteristics of the i-th link at time t. The k-th normalized index, Indicates the number of normalization indicators. The weight of the k-th normalized metric can be configured based on task requirements, reliability, latency, and packet delivery rate. The instantaneous link quality index reflects the physical layer quality of the link at that moment.

[0039] Step 2: Based on the historical features of the link and the link status label, train a logistic regression model using the maximum likelihood estimation method, and input the real-time features of the link into the trained logistic regression model to predict the quality index of the disconnection risk.

[0040] In a preferred but non-limiting embodiment of the present invention, step 2 includes: Step 2.1: Collect the historical characteristics and corresponding link status labels of each link when it is in normal communication state and when it is disconnected, and construct a historical sample set.

[0041] Link status labels are used to characterize whether a communication interruption has occurred within the time window.

[0042] Step 2.2: Establish a logistic regression model based on the historical sample set, and train the logistic regression model using the maximum likelihood estimation method.

[0043] More preferably, step 2.2 includes: Step 2.2.1: Collect historical characteristics and link status labels of each link when it is in normal communication state and when it is disconnected to establish a logistic regression model. The logistic regression model is used to determine the link availability probability of each training sample in the training set, expressed by the following formula: (5) In the formula, Indicates link historical characteristics The probability that the downlink is available. Represents link state variables, when =1 indicates that the link is in an available state. A value of 0 indicates that the link has dropped. It represents the historical characteristics of the link, including indicators such as normalized received signal strength, signal-to-noise ratio, packet reception rate, and round-trip delay. This represents the Sigmoid function, used to map the result of a linear combination to the interval [0, 1] to represent the probability that the link is in an available state. This is a bias term used to characterize the basic availability level of the link in the absence of feature input. , , and This represents the link feature weight, used to characterize the degree of impact of different link features on the risk of link disconnection. Its positive or negative sign and numerical value reflect the promoting or inhibiting effect of the corresponding indicator on link stability. and , and These represent the historical signal-to-noise ratio, historical packet reception rate, historical received signal strength, and historical round-trip delay, respectively, after normalization.

[0044] The core idea of ​​this invention is to estimate the probability distribution of whether a link remains available or goes offline given the current link characteristics by learning the statistical relationship between link features and link status in historical samples.

[0045] Step 2.2.2: Based on the historical sample set and the logistic regression model, train the logistic regression model through maximum likelihood estimation, find a set of optimal logistic regression parameters, and obtain the trained logistic regression model.

[0046] The set of logistic regression parameters The solution is obtained using the maximum likelihood estimation method, which involves finding a set of parameters in the historical sample set that maximizes the joint probability of observing current link state samples under these parameters. In this way, the model can adaptively learn the contribution relationship between different link characteristics and the risk of connection drops in real-world operating environments.

[0047] Step 2.3, real-time collected link features Input the trained logistic regression model to determine the probability of link availability in the current state, and set it as the quality index of disconnection risk. It can be expressed by the following formula: (6) Step 3: Determine short-term disturbance features based on the dynamic convolution kernel weights of the power communication network environment characteristics, determine trend features based on the short-term disturbance features, fuse the trend features and short-term disturbance features through cross-scale gating fusion method, input the fusion result into encoder-decoder to generate link quality prediction values, and aggregate the link quality prediction values ​​into link trend quality index.

[0048] In a preferred but non-limiting embodiment of the present invention, step 3 includes: Step 3.1: Obtain link quality characteristics within the historical time window. It can be expressed by the following formula: (7) In the formula, Indicates in Normalized received signal strength at any given time. Indicates in Normalized signal-to-noise ratio at any given moment. Indicates in Normalized packet reception rate at any given time. Indicates in The normalized round-trip time at any given moment, along with the above metrics, collectively reflects the communication status of the link at different time scales. Indicates the length of the historical time window. An index representing historical time.

[0049] Step 3.2: To address the issue of insufficient response capability of the original Autoformer to high-frequency disturbances, this invention introduces short-term disturbance enhancement at the input end. Dynamic convolution kernel weights are selected according to the different environmental characteristics of the power communication network. Within each sliding time window of the link quality feature, short-term disturbance features are extracted through dynamic convolution kernel weights. The extracted short-term disturbance features are added to the original input in a residual manner to form the enhanced link quality features, thus avoiding damage to the overall structure of the original sequence.

[0050] More preferably, step 3.2 includes: Step 3.2.1: Obtain different environmental characteristics of the power communication network, including environmental stability index, environmental dynamic index and environmental interference index. Among them, the environmental stability index is the variance of the link quality characteristics within the historical time window, the environmental dynamic index is the average value of the change rate of the feature vector of the link quality characteristics at adjacent time points within the historical time window, and the environmental interference index is the fluctuation rate of the signal-to-noise ratio in the link quality characteristics within the historical time window.

[0051] Step 3.2.2 assumes that the environmental characteristics follow a mixture distribution consisting of M Gaussian distributions, where M=4 corresponds to four typical environmental categories. Based on the different environmental characteristics of the power communication network, the environmental probability density function of the power communication network is constructed using a Gaussian Mixture Model (GMM), expressed by the following formula: (8) In the formula, Indicates known model parameters Under these conditions, environmental characteristics were observed. The probability density value, This represents the environmental characteristics at time t. This represents the set of model parameters, containing the mixing coefficients of all Gaussian components. Mean vector Covariance Matrix , This represents the mixing coefficient for the m-th environmental category. Let represent the mean vector of the m-th environment category. Let m be the covariance matrix of the m-th environmental category. This means that given a mean vector Covariance Matrix Under these conditions, environmental characteristics The probability density of the multivariate Gaussian distribution.

[0052] M represents the number of environmental categories. When m=1, it represents a stable indoor equipment communication scenario, such as, but not limited to, fixed wireless communication equipment such as fixed monitoring terminals in substation control rooms and cable trench monitoring equipment. There is less external interference and the communication link quality is stable. When m=2, it represents a periodic interference scenario in urban power distribution networks, such as, but not limited to, urban ring main unit monitoring terminals, distributed photovoltaic inverter communication modules, and electric vehicle charging pile communication units. There is regular electromagnetic interference, and the communication quality fluctuates periodically. When m=3, it represents a mobile inspection and equipment operation scenario, such as, but not limited to, communication between inspection drones and base stations, and communication terminals of mobile repair vehicles. The communication terminals move rapidly, and the frequent changes in equipment position lead to rapid signal attenuation and frequent changes in communication quality. When m=4, it represents a severe weather interference scenario, such as transmission line monitoring during thunderstorms and power distribution equipment communication during rain and snow. Severe weather absorbs and attenuates radio signals, and lightning generates strong electromagnetic pulse interference, causing a sharp deterioration in the communication link quality.

[0053] Step 3.2.3: Based on the environmental probability density function, the optimal parameters of the environmental probability density function are obtained by training using the Expectation-Maximization (EM) algorithm.

[0054] Step 3.2.4: Calculate the probability density value for each environmental category based on the optimal parameters, and sum the weighted probability density values ​​for each environmental category to determine the weighted sum of the probability densities for all environmental categories, expressed by the following formula: (9) In the formula, This represents the total probability density of environmental feature E(t) across the entire Gaussian mixture model. This represents the optimal mixing coefficient for the m-th environmental category. Let represent the optimal mean vector for the m-th environment category. Let represent the optimal covariance matrix for the m-th environmental category.

[0055] Step 3.2.5: Calculate the posterior probability of the current environment belonging to each category by summing the weighted probability densities of all environment categories using Bayes' theorem, expressed as follows: (10) In the formula, Let represent the posterior probability that the current environment belongs to class m, and satisfy . , Step 3.2.6: Calculate the dynamic convolution kernel weights based on the posterior probability of the current environment belonging to each category, expressed by the following formula: (11) In the formula, Indicates the dynamic convolution kernel weights. Indicates the baseline convolution kernel weights. This represents the adjustment weight for the m-th environment category.

[0056] Within each sliding time window of the link quality features, short-term perturbation features are extracted using dynamic convolutional kernel weights, expressed as follows: (12) In the formula, This represents the short-term disturbance characteristics extracted at time t, used to reflect the instantaneous fading, sudden interference, or rapid fluctuation state of the link. Indicates the dynamic convolution kernel weights. This indicates that time t is the endpoint and the length is... Link quality characteristics collected within the time window. This represents the perturbation bias term.

[0057] The extracted short-term perturbation features are added to the original input as residuals to form the enhanced link quality features, avoiding damage to the overall structure of the original sequence, as expressed by the following formula: (13) In the formula, This indicates the link quality characteristics enhanced by a short-term disturbance.

[0058] It is worth noting that this invention solves the problem of insufficient modeling ability of traditional time series prediction models for short-term sudden disturbances in power communication links by using environment-adaptive classification based on Gaussian mixture models and environment-adaptive convolution kernel parameter adjustment strategies. It overcomes the limitation that fixed model parameters cannot adapt to complex and changing operating environments, improves the model's adaptability to different operating environments (stable / dynamic / harsh), reduces prediction errors caused by environmental changes, and improves the accuracy and adaptability of link quality trend prediction.

[0059] Step 3.3: The enhanced link quality features are decomposed into trend vector, period vector, and perturbation vector using the Autoformer sequence, as expressed by the following formula: (14) In the formula, T(t) represents the trend vector, which reflects the upward or downward trend of link quality over a long time scale; S(t) represents the period vector, which reflects the periodic change pattern in the link caused by environmental or service cycles; and D(t) represents the disturbance vector, which describes the high-frequency fluctuations in the link caused by instantaneous interference, sudden blockage, or noise.

[0060] Step 3.4: Extract trend features from the trend vector using an encoder. Trend features and short-term perturbation features are concatenated, and the cross-scale fusion weights of the two features in the fusion result are dynamically adjusted by a learnable gating factor, as expressed by the following formula: (15) In the formula, The cross-scale fusion weights calculated at time t represent the proportion of short-term perturbation features in the current fusion result. The learnable gating factor includes the learnable weight matrix. and learnable bias vector .

[0061] Step 3.5: Based on the cross-scale fusion weights, the features at different scales are weighted and reconstructed to obtain the cross-scale fusion features, expressed by the following formula: (16) In the formula, This represents the cross-scale fusion feature at time t.

[0062] When the link is under severe fluctuation or strong interference, cross-scale fusion weights Automatic scaling makes the model pay more attention to short-term disturbance characteristics; when the link is in a relatively stable state, trend characteristics dominate, thus ensuring the smoothness and stability of the prediction results.

[0063] Step 3.6: Based on the Autoformer encoder-decoder, the encoder utilizes an autocorrelation-based attention mechanism to fuse features across scales. The encoder output is then passed through multiple layers of autocorrelation and cross-attention in the decoder to generate future link quality predictions for a set of R time points. ,in, Indicates future time The predicted link quality is expressed by the following formula: (17) In the formula, This represents the predicted normalized received signal strength at time t+R. This represents the predicted normalized signal-to-noise ratio at time t+R. This represents the predicted normalized packet reception rate at time t+R. This represents the predicted normalized round-trip time at time t+R.

[0064] Step 3.7: Aggregate the predicted link quality values ​​into a link trend quality index, expressed by the following formula: (18) In the formula, The link trend quality index represents the link i-th link, reflecting the overall evolution trend of the link over future time scales. Its results can serve as an important decision-making basis for preventing the continuous deterioration of link quality in advance in multi-link switching strategies. This invention represents a preset mapping function used to comprehensively characterize the future stability of the link. To obtain the weighted average, the following formula is used: (19) In the formula, , , and Here, R represents the weighting coefficient, and R represents the total number of future moments to be predicted.

[0065] It is worth noting that this invention improves the traditional Autoformer encoder-decoder model by enhancing short-term perturbations and cross-scale fusion, avoiding prediction bias caused by single-scale feature extraction, solving the problem of co-modeling long-term trends and short-term perturbations, improving the prediction model's ability to capture features at multiple time scales, and reducing mispredictions caused by short-term fluctuations.

[0066] Step 4: Calculate the fusion weights with the switching decision accuracy as the objective. Based on the fusion weights, merge the instantaneous link quality index, the disconnection risk quality index, and the link trend quality index into a comprehensive link quality, expressed by the following formula: (20) In the formula, Indicates the overall link quality. This represents the instantaneous link quality index of the i-th link at time t; This represents the quality index of the risk of the i-th link dropping at time t; The average link trend quality index of the i-th link within a future time window is expressed by the following formula: (twenty one) In the formula, The value represents the length of the time window for predicting future trends, and h represents the index of the prediction step.

[0067] Fusion weight parameters , and The following constraints must be met: (twenty two) To ensure consistency between the comprehensive link quality evaluation results and the actual handover performance, this invention adjusts the fusion weight parameters based on historical sample data. , and The optimization is determined by the following steps. Specifically, based on a dataset containing Ns historical samples, the optimization problem is constructed with the goal of maximizing the accuracy of switching decisions: (twenty three) In the formula, Indicates the number of historical samples. This represents the overall link quality value corresponding to the i-th link in the n-th historical sample. This represents the identifier of the actual optimal handover link in the nth sample. This indicates the indicator function. It optimizes using historical data to maximize the matching rate between the link with the highest overall score and the actual optimal link.

[0068] It is worth noting that this invention solves the problem of one-sidedness in single-dimensional quality assessment by fusing three-dimensional quality indices of instantaneous quality, disconnection risk, and future trends, and optimizing multi-objective weights based on historical data. It also addresses the issue of fixed weights failing to adapt to changes in network conditions, overcoming the reliance on experience and subjectivity in manually setting fusion weights. Furthermore, it resolves the challenge of balancing and compromising between multiple quality dimensions, improving the system's adaptability under different network conditions, increasing handover success rate and service quality, reducing decision-making errors caused by a single assessment dimension, minimizing performance fluctuations due to subjective weight settings, reducing the randomness and uncertainty of handover decisions, and lowering the handover failure rate and the probability of service interruption.

[0069] Step 5: If the link switching conditions are met, determine the link with the highest comprehensive score based on the overall link quality. If the link with the highest comprehensive score is not the link that was switched last time, switch the current link to the link with the highest comprehensive score. If the link with the highest comprehensive score is the link that was switched last time, switch the current link to the link with the highest comprehensive score when the score difference between the current link and the link with the highest comprehensive score is greater than a set threshold. If the overall link quality of the new link after switching is continuously lower than the difference between the overall link quality of the original link before switching and the fallback threshold within a set period of time, switch back to the original link.

[0070] In a preferred but non-limiting embodiment of the present invention, step 5 includes: Step 5.1, if the disconnection risk quality index Exceeding the preset threshold If the current link is determined to be in a high-risk state, a link switch is required; if the trend quality index for the next N steps... The minimum value is lower than the threshold. If the current link is determined to deteriorate rapidly in the short term, a link switch is required. If either of these two conditions is met, the current link is deemed unsuitable for continued use, and the link switchover process begins.

[0071] Step 5.2: Evaluate the overall score of the link by considering link quality, link delay, and communication cost, expressed by the following formula: (twenty four) In the formula, This represents the overall score of the i-th link. , and These represent the weighting coefficients for quality, delay, and cost, respectively. This represents the latency metric of the i-th link, used to indicate the priority of low-latency links. This represents the communication cost of the i-th link. The communication cost includes, but is not limited to, eSIM, traffic consumption, and link power consumption, and is used to achieve an economic balance among multiple available links.

[0072] Select the link with the highest overall score It can be expressed by the following formula: (25) Step 5.3: If the link with the highest overall score was the link in the last switch, calculate the score difference between the current link and the link with the highest overall score. If the score difference is greater than a set threshold, perform a link switch; otherwise, do not switch. The link in the last switch specifically refers to the link that the system was using before switching to the current link in the most recent successfully executed switch operation.

[0073] After identifying the target link, it is necessary to ensure that the handover process does not cause oscillations or frequent switchbacks. This invention employs a jitter suppression mechanism combining a minimum hold time and a switchback threshold to avoid system instability caused by frequent handovers. The minimum hold time ensures that the current link will not immediately trigger a handover due to short-term fluctuations within a certain time window; while the switchback threshold prevents a switchover from immediately reverting to the previous one.

[0074] After the switching trigger conditions are met, the comprehensive score of each candidate link is calculated. If the link with the highest score is a link that has not been used recently, then the link is switched to directly. If the link has been switched away from recently, the switch is only performed when its score is higher than a certain threshold of the current link. After the switch, if the quality of the new link is lower than the quality of the original link minus the fallback threshold during the monitoring period, the failure fallback is performed.

[0075] In actual execution, the system first caches data to ensure that currently unsent data is not lost; then it switches the link interface; and according to the specific communication protocol, it completes the network re-registration or link reconnection process, while refreshing the routing table to ensure that the data flow is correctly forwarded to the new link path. This step guarantees the atomicity of the switching action and the synchronization of the data plane and control plane.

[0076] Step 5.4: To avoid communication degradation after switching to a certain link, this invention designs a failure fallback mechanism based on reverse link quality detection. After the switch, the system will continuously evaluate the overall link quality of the new link. If its quality is lower than that of the old integrated link within the set time period Subtracting a backoff threshold δ satisfies: (26) In the formula, δ represents the backoff threshold. In order to avoid the system repeatedly switching between two links due to normal small fluctuations in link quality or measurement noise, the system requires that the quality of the new link must be significantly and continuously better than that of the old link in order to consider the handover successful. The backoff threshold δ is a specific numerical threshold that defines the quality of the new link must be significantly and continuously better than that of the old link. In this invention, the backoff threshold δ is set to 0.05.

[0077] If the handover fails, the system will automatically fall back to the previous link. This mechanism enables the system to self-correct, preventing communication quality degradation due to prediction errors or short-term sudden interference.

[0078] The quality of the new link must not only be worse than that of the old link, but must be worse than the fallback threshold δ for the system to determine that the switchover has failed and execute the fallback. If the quality of the new link is only slightly lower than that of the old link (the difference is less than δ), the system will consider this a normal fluctuation and choose to continue to observe and tolerate it without triggering the fallback.

[0079] It is worth noting that this invention, through a dual-trigger condition judgment mechanism, comprehensively calculates quality, latency, and cost scores, uses a quality threshold difference based on the overall link quality for anti-jitter, and employs a quality monitoring and failure fallback mechanism. This solves the problem of delayed response in traditional handover methods, overcomes link oscillations and instability caused by frequent handovers, avoids the limitation of handover decisions that only consider quality while ignoring cost-effectiveness, overcomes the defect of fixed parameters being unable to adapt to dynamic network changes, improves the timeliness of handover decisions, enhances the stability and reliability of the handover process, improves the economy of system resource utilization, reduces average handover latency and service interruption time, reduces the number of unnecessary handovers and link oscillation frequency, reduces communication costs and energy consumption, and reduces service losses caused by handover failures.

[0080] A multi-link selection strategy based on comprehensive scoring and an intelligent handover process with minimum hold time, handover threshold and failure fallback mechanism are proposed, which significantly improves the communication continuity, handover accuracy and stability of power terminals and is suitable for promotion and application in multi-link access equipment of new power systems.

[0081] In summary, after completing the multi-link quality assessment, this invention achieves intelligent multi-link handover in power scenarios through an intelligent decision-making mechanism consisting of "handover triggering, candidate link selection, handover execution and jitter suppression, and failure fallback." This mechanism not only responds quickly to risks and avoids link deterioration in advance, but also ensures the stability and robustness of the handover process and has automatic recovery capabilities in the event of handover failure, thereby significantly improving the continuity and reliability of power communication systems.

[0082] Embodiment 2 of the present invention provides a multi-link handover system for power communication networks based on link quality prediction, which implements a multi-link handover method for power communication networks based on link quality prediction provided in Embodiment 1, including: The link instantaneous quality index calculation module is used to calculate the link instantaneous quality index based on the link's real-time characteristics through weighted normalization. The disconnection risk quality index calculation module is used to train a logistic regression model based on historical link features and link status labels using the maximum likelihood estimation method, and input real-time link features into the trained logistic regression model to predict the disconnection risk quality index. The link trend quality index calculation module is used to determine short-term disturbance features based on the dynamic convolution kernel weights based on the characteristics of the power communication network environment, determine trend features based on the short-term disturbance features, fuse the trend features and short-term disturbance features through a cross-scale gating fusion method, input the fusion result into the encoder-decoder to generate link quality prediction values, and aggregate the link quality prediction values ​​into the link trend quality index. The fusion module is used to combine the instantaneous link quality index, the disconnection risk quality index, and the link trend quality index into a comprehensive link quality index. The switching module is used to determine the link with the highest comprehensive score based on the overall link quality if the switching conditions are met. If the link with the highest comprehensive score is not the link that was switched in the last time, the current link will be switched to the link with the highest comprehensive score. If the link with the highest comprehensive score is the link that was switched in the last time, the current link will be switched to the link with the highest comprehensive score when the score difference between the current link and the link with the highest comprehensive score is greater than a set threshold. If the overall link quality of the new link after switching is lower than the difference between the overall link quality of the original link before switching and the fallback threshold within a set period of time, the original link will be switched back.

[0083] This system can be integrated into the following typical power terminals: This invention includes a multi-wireless interface board for substation intelligent monitoring terminals; a ZigBee / LoRa / 4G multi-mode communication module for distribution zone concentrators; a wireless backup link for distribution automation terminals (FTU / DTU); and automatic link switching for low-power wireless e-ink terminals. In actual power field applications, this invention can predict link deterioration trends 2–10 seconds in advance, improving switching accuracy by 12–25% and reducing false switching by more than 30%, significantly improving communication reliability. If the conditions remain in effect, it automatically reverts to the old link.

[0084] Compared with the prior art, the beneficial effects of the present invention include at least the following: This application addresses heterogeneous multi-link access scenarios in power communication networks, including ZigBee, LoRa, NB-IoT, and 4G / 5G. It constructs a three-layer link quality assessment system consisting of instantaneous link quality assessment, dropout risk prediction, and future trend prediction. This system not only quantifies the current state of the link to achieve comparability between different communication technologies, but also uses a predictive model to pre-determine the future evolution trend and dropout probability of the link. The multi-layer assessment results are integrated to form a unified link quality index, which is used to directly guide multi-link selection and switching decisions, enabling comparability evaluation between links of different communication standards and avoiding switching decision bias caused by differences in the indicator system. This invention employs hierarchical modeling and fusion evaluation of instantaneous link quality, downtime risk, and future evolution trends. This enables link quality assessment to no longer rely solely on a single moment or indicator, but to simultaneously reflect the current state of the link, potential failure risks, and future trends, thereby achieving early detection and proactive avoidance of link degradation. Furthermore, by constructing an intelligent decision-making mechanism that includes handover triggering, candidate link selection, handover execution, jitter suppression, and failure fallback, the probability of erroneous and frequent handovers is effectively reduced, improving the stability and robustness of multi-link handover processes and significantly enhancing the continuity and reliability of power communication systems in complex field environments. This application introduces a drop risk prediction based on historical samples before link switching decisions to quantify the probability of a link dropping in the future. By predicting the drop risk and future trends, it enables early switching before link deterioration, thereby reducing the probability of communication interruption. Compared with the scheme that only performs parameter estimation, this application can significantly improve the stability and communication continuity of multi-link access in complex power field environments. By predicting future link quality trends through multi-scale sequence decomposition and improvements to the Autoformer encoder-decoder, a three-layer link quality assessment mechanism covering the current state, potential risks, and future trends is constructed. Based on the fused unified link quality index, prediction-driven intelligent multi-link switching is achieved. By predicting the future evolution trend of links, link switching is transformed from a reactive response to a proactive avoidance, effectively reducing the risk of communication interruption caused by sudden link deterioration. By fusing the results of multi-layer quality assessment, the stability and accuracy of switching decisions are improved in scenarios with frequent fluctuations in multiple links. This application can reduce unnecessary frequent switching, reduce system resource consumption, and improve the overall reliability of power communication networks. This invention is applicable to the access control of power Internet of Things (IoT) devices such as smart meters, power distribution terminals, and low-power monitoring devices between heterogeneous wireless links such as ZigBee, LoRa, NB-IoT, and 4G / 5G.

[0085] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A multi-link handover method for power communication networks based on link quality prediction, characterized in that, include: Based on the real-time characteristics of the link, the instantaneous quality index of the link is calculated by weighted normalization; Based on historical link features and link status labels, a logistic regression model is trained using the maximum likelihood estimation method. Real-time link features are then input into the trained logistic regression model to predict the quality index of disconnection risk. Short-term disturbance features are determined based on dynamic convolutional kernel weights based on the characteristics of the power communication network environment. Trend features are determined based on short-term disturbance features. Trend features and short-term disturbance features are fused using a cross-scale gating fusion method. The fusion result is input into an encoder-decoder to generate link quality prediction values. The link quality prediction values ​​are aggregated into a link trend quality index. The instantaneous link quality index, the disconnection risk quality index, and the link trend quality index are combined into a comprehensive link quality index. If the link switching conditions are met, the link with the highest comprehensive score is determined based on the overall link quality. If the link with the highest comprehensive score is not the link that was switched in the last time, the current link is switched to the link with the highest comprehensive score. If the link with the highest comprehensive score is the link that was switched in the last time, the current link is switched to the link with the highest comprehensive score when the score difference between the current link and the link with the highest comprehensive score is greater than a set threshold. If the overall link quality of the new link after switching is continuously lower than the difference between the overall link quality of the original link before switching and the fallback threshold within a set period of time, the original link is switched back.

2. The multi-link handover method for power communication networks based on link quality prediction according to claim 1, characterized in that: The instantaneous link quality index is expressed by the following formula: In the formula, This represents the instantaneous quality index of the i-th link, where i=1 represents a ZigBee link, i=2 represents a LoRa link, and i=3 represents a 4G link. This represents the real-time characteristics of the i-th link at time t. The k-th normalized index, Indicates the number of normalization indicators. It represents the weight of the k-th normalized index.

3. The multi-link handover method for power communication networks based on link quality prediction according to claim 1, characterized in that: The predicted quality index for disconnection risk includes: Collect historical characteristics and link status labels of each link when it is in normal communication state and when it is disconnected to establish a logistic regression model, and train the logistic regression model by the maximum likelihood estimation method. The real-time collected link features are input into the trained logistic regression model to determine the probability that the link is available in the current state, which is set as the disconnection risk quality index.

4. The multi-link handover method for power communication networks based on link quality prediction according to claim 3, characterized in that: Establish a logistic regression model, expressed by the following formula: In the formula, Indicates link historical characteristics Probability of downlink availability, link history characteristics Includes normalized historical signal-to-noise ratio metrics Historical packet reception rate indicators Historical received signal strength indicators Historical round-trip delay indicators , Indicates the link status label, when =1 indicates that the link is in an available state. A value of 0 indicates that the link has dropped. This represents the Sigmoid function. For bias terms, , , and This represents the link feature weights.

5. The multi-link handover method for power communication networks based on link quality prediction according to claim 1, characterized in that: The calculation of dynamic convolution kernel weights for power communication network environmental characteristics includes: Based on the different environmental characteristics of power communication networks, an environmental probability density function of power communication networks is constructed using a Gaussian mixture model. The optimal parameters of the environmental probability density function are obtained by training the function using the expectation-maximization algorithm. Calculate the probability density value of each environmental category based on the optimal parameters, and sum the weighted probability density values ​​of each environmental category to determine the weighted probability density sum of all environmental categories. The posterior probability of the current environment belonging to each category is calculated by summing the weighted probability densities of all environment categories based on Bayes' theorem. The dynamic convolution kernel weights are calculated based on the posterior probability that the current environment belongs to each category.

6. The multi-link handover method for power communication networks based on link quality prediction according to claim 5, characterized in that: Determining the characteristics of short-term disturbances includes: The link quality features within the historical time window are obtained. Within each sliding time window of the link quality features, short-term perturbation features are extracted using dynamic convolutional kernel weights, as expressed by the following formula: In the formula, This represents the short-term perturbation features extracted at time t. Indicates the dynamic convolution kernel weights. This indicates that time t is the endpoint and the length is... Link quality characteristics collected within the time window. This represents the perturbation bias term.

7. The multi-link handover method for power communication networks based on link quality prediction according to claim 1, characterized in that: The generated link quality prediction values ​​include: The short-term perturbation features are added to the original input as residuals to form the enhanced link quality features; The enhanced link quality features are decomposed into a sequence to obtain a trend vector; The trend vector is processed by an encoder to extract trend features. The trend features and short-term perturbation features are concatenated and the cross-scale fusion weights of the two features in the fusion result are dynamically adjusted by a learnable gating factor. Based on the cross-scale fusion weights, features of different scales are weighted and reconstructed to obtain cross-scale fusion features; The encoder uses an autocorrelation attention mechanism to encode cross-scale fused features, and the encoder output is passed through the decoder's multi-layer autocorrelation attention and cross attention to generate link quality prediction values.

8. A multi-link handover method for power communication networks based on link quality prediction according to claim 7, characterized in that: Cross-scale fusion features are expressed by the following formula: In the formula, This represents the cross-scale fusion feature at time t. Indicates trend characteristics, This represents the short-term perturbation features extracted at time t. The cross-scale fusion weights calculated at time t are represented by the following formula: Learnable gating factors include learnable weight matrices. and learnable bias vector , This represents the Sigmoid function.

9. A multi-link handover method for power communication networks based on link quality prediction according to claim 1, characterized in that: The link with the highest overall score is determined based on the overall link quality, expressed by the following formula: In the formula, This indicates the link with the highest overall score. The overall score of the i-th link is represented by the following formula: In the formula, , and These represent the weighting coefficients for quality, delay, and cost, respectively. To improve overall link quality, This represents the latency metric for the i-th link. This represents the communication cost of the i-th link.

10. A multi-link handover system for power communication networks based on link quality prediction, comprising a multi-link handover method for power communication networks based on link quality prediction according to any one of claims 1-9, characterized in that: The link instantaneous quality index calculation module is used to calculate the link instantaneous quality index based on the link's real-time characteristics through weighted normalization. The disconnection risk quality index calculation module is used to train a logistic regression model based on historical link features and link status labels using the maximum likelihood estimation method, and input real-time link features into the trained logistic regression model to predict the disconnection risk quality index. The link trend quality index calculation module is used to determine short-term disturbance features based on the dynamic convolution kernel weights based on the characteristics of the power communication network environment, determine trend features based on the short-term disturbance features, fuse the trend features and short-term disturbance features through a cross-scale gating fusion method, input the fusion result into the encoder-decoder to generate link quality prediction values, and aggregate the link quality prediction values ​​into the link trend quality index. The fusion module is used to combine the instantaneous link quality index, the disconnection risk quality index, and the link trend quality index into a comprehensive link quality index. The switching module is used to determine the link with the highest comprehensive score based on the overall link quality if the switching conditions are met. If the link with the highest comprehensive score is not the link that was switched in the last time, the current link will be switched to the link with the highest comprehensive score. If the link with the highest comprehensive score is the link that was switched in the last time, the current link will be switched to the link with the highest comprehensive score when the score difference between the current link and the link with the highest comprehensive score is greater than a set threshold. If the overall link quality of the new link after switching is lower than the difference between the overall link quality of the original link before switching and the fallback threshold within a set period of time, the original link will be switched back.

Images