Smart park multi-voltage-level direct current carrier adaptive channel modeling method and system

Abstract

The invention relates to the technical field of electronic power, and particularly discloses a smart park multi-voltage-class direct-current carrier self-adaptive channel modeling method and system, and the method comprises the steps: collecting direct-current carrier channel data in a smart park, the direct-current carrier channel data comprising multi-voltage-class basic data and distributed power supply data; and extracting four-path channel characteristic parameters of the direct-current carrier channel data to form a characteristic set. According to the invention, the method achieves the precise and real-time modeling and adaptive tracking of the complex DC carrier channel of the smart park through the construction of an innovative four-path channel model and the combination of an edge-cloud collaborative interpretable neural network XNN architecture, guarantees the highest robustness and high availability of a communication link under the complex working condition of a power system, and improves the reliability of the communication link. The problems that in a traditional scheme, channel models are mismatched, interference sources are unknown, and the system can only be passively adapted are solved.

Description

Technical Field

[0001] This invention relates to the field of electronic power technology, and in particular to a method and system for adaptive channel modeling of DC carriers at multiple voltage levels in smart parks. Background Technology

[0002] In smart parks, DC distribution networks are widely used due to their ease of integration with distributed generation systems (DERs) such as photovoltaics and energy storage. DC-PLC technology, utilizing existing power lines for communication, is a preferred solution for achieving intelligent management, energy dispatch, and equipment monitoring within the park. However, traditional PLC channel modeling methods are mostly designed for AC or stable DC environments, resulting in relatively fixed model structures. In the complex DC environment of smart parks, multiple voltage levels such as 10kV, 380V, and 220V are connected to the grid, and numerous power electronic devices such as inverters and converters are connected to the grid as DER interfaces. These devices are not only loads but also dynamically changing noise and reflection sources, introducing interference paths to the DC-PLC channel that traditional models have not adequately considered.

[0003] Existing channel models generally suffer from two major drawbacks: First, poor adaptability. The power fluctuations and interface impedance changes of DERs are highly time-varying and random, causing the channel characteristics to change rapidly. Static channel models cannot track these changes in real time, resulting in high bit error rates and insufficient reliability.

[0004] Second, there is a lack of interpretability. When communication quality degrades, traditional models cannot diagnose the dominant source of interference, such as load reflection or DER interference. This forces the system to passively retransmit or slow down at the communication layer, failing to proactively suppress interference at the root cause, i.e., the physical layer. Furthermore, it cannot achieve collaborative optimization between the communication system and the energy management system (EMS), resulting in the robustness of the communication link being limited by the operating status of the power system. Summary of the Invention

[0005] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, the purpose of this invention is to propose a method and system for adaptive channel modeling of DC carriers at multiple voltage levels in smart parks, so as to achieve accurate and real-time modeling of DC carrier channels.

[0006] To achieve the above objectives, a first aspect of the present invention proposes a method for adaptive channel modeling of multi-voltage level DC carriers in smart parks, the method comprising:

[0007] Collect DC carrier channel data within the smart park. The DC carrier channel data includes basic data for multiple voltage levels and distributed power source data. The multiple voltage levels include 10kV, 380V, and 220V voltage levels, and the distributed power sources include photovoltaic power sources or energy storage power sources.

[0008] The four-path channel feature parameters of the DC carrier channel data are extracted to form a feature set; wherein, the four paths include the direct transmission path, the load reflection path, the environmental interference path, and the distributed power source reflection path;

[0009] Based on the feature set, an interpretable neural network XNN channel model is constructed and applied through an edge-cloud collaborative approach. The XNN channel model has four sub-networks corresponding to the four paths respectively. The construction and application include: training the XNN channel model on the cloud platform, compressing the trained model and distributing it to the edge nodes, and the edge nodes loading the compressed model to perform channel modeling inference.

[0010] The fluctuations in the distributed power source data are monitored, and when the fluctuations meet preset adjustment conditions, the parameters of the XNN channel model are dynamically adjusted to achieve adaptive updating of the channel model.

[0011] To achieve the above objectives, a second aspect of the present invention proposes a smart campus multi-voltage level DC carrier adaptive channel modeling system, comprising:

[0012] The data acquisition module is used to collect DC carrier channel data within the smart park. The DC carrier channel data includes basic data for multiple voltage levels and distributed power source data. The multiple voltage levels include 10kV, 380V, and 220V voltage levels, and the distributed power sources include photovoltaic power sources or energy storage power sources.

[0013] The feature extraction module is used to extract four-path channel feature parameters based on the DC carrier channel data to form a feature set; wherein, the four paths include direct transmission path, load reflection path, environmental interference path, and distributed power source reflection path;

[0014] The channel modeling module is used to construct and apply an interpretable neural network XNN channel model based on the feature set through an edge-cloud collaborative approach. The XNN channel model has four sub-networks corresponding to the four paths respectively. Specifically, the channel modeling module is used to: train the XNN channel model on the cloud platform, compress the trained model and distribute it to the edge nodes, and the edge nodes load the compressed model to perform channel modeling inference.

[0015] An adaptive adjustment module is used to monitor the fluctuations of the distributed power source data. When the fluctuations meet preset adjustment conditions, the parameters of the XNN channel model are dynamically adjusted to achieve adaptive updates of the channel model.

[0016] To achieve the above objectives, a third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described smart park multi-voltage level DC carrier adaptive channel modeling method.

[0017] The smart park multi-voltage level DC carrier adaptive channel modeling method and system of this invention realizes accurate, real-time modeling and adaptive tracking of complex DC carrier channels in smart parks by constructing an innovative four-path channel model and combining it with the edge-cloud collaborative interpretable neural network XNN architecture.

[0018] The core advantages of this solution lie in its interpretability and proactive closed-loop mechanism: First, the lightweight XNN engine deployed at the edge can not only predict channel conditions but also diagnose the dominant interference paths that lead to communication quality degradation in real time. Second, based on this interpretable diagnostic result, the system can proactively initiate preventative closed-loop optimization by sending precise control commands to the distributed power supply or load scheduling modules of the power physical layer to suppress interference at its source. Furthermore, when physical layer control is obstructed due to priority conflicts, the adaptive avoidance strategy of this solution at the communication layer can utilize the prediction results of XNN to proactively adjust the parameters of the communication transceiver in a feedforward manner, ensuring the highest robustness and high availability of the communication link under complex power system operating conditions, and solving the problems of channel model mismatch, unclear interference sources, and the communication system's passive adaptation in traditional solutions. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the smart park multi-voltage level DC carrier adaptive channel modeling method provided by the present invention.

[0020] Figure 2 This is a schematic diagram of the four-path DC carrier channel model in the smart park multi-voltage level DC carrier adaptive channel modeling method provided by the present invention;

[0021] Figure 3 This is a waterfall diagram of the real-time output components of the four sub-networks of XNN in the smart park multi-voltage level DC carrier adaptive channel modeling method provided by the present invention;

[0022] Figure 4 This is a schematic diagram of the normalized trajectory of distributed power source reflection path characteristics in the smart park multi-voltage level DC carrier adaptive channel modeling method provided by the present invention;

[0023] Figure 5 The nonlinear adaptive adjustment factor in the smart park multi-voltage level DC carrier adaptive channel modeling method provided by this invention is... A schematic diagram of the S-curve;

[0024] Figure 6 This is a schematic diagram of the signal-to-noise ratio rolling prediction with a prediction time window of 15 minutes in the smart park multi-voltage level DC carrier adaptive channel modeling method provided by the present invention;

[0025] Figure 7 This is a schematic diagram comparing the BER simulation of the three strategies under high interference conditions in the smart park multi-voltage level DC carrier adaptive channel modeling method provided by this invention.

[0026] Figure 8 This is a schematic diagram of the structure of the smart park multi-voltage level DC carrier adaptive channel modeling system provided by the present invention;

[0027] Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0028] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0029] The following description, with reference to the accompanying drawings, illustrates a smart park multi-voltage level DC carrier adaptive channel modeling method, system, and electronic device according to embodiments of the present invention.

[0030] Example 1:

[0031] This embodiment details the basic framework of a multi-voltage-level DC carrier adaptive channel modeling method for smart parks. The purpose of this embodiment is to provide a basic platform for channel modeling with edge-cloud collaboration capabilities and deep interpretability.

[0032] Before describing the steps of this method in detail, a preferred system architecture for implementing the method is first described. This invention is preferably implemented in an edge-cloud collaborative distributed computing architecture. Specifically, it includes:

[0033] Cloud Platform: The cloud platform is the central hub for model training and global optimization in this invention, possessing powerful computing capabilities and massive storage resources. Its core responsibility is to aggregate channel feature data and measured label data from various edge nodes within the campus, utilizing this data for deep training, validation, and optimization of complex and computationally intensive interpretable neural network (XNN) models. After training, the cloud platform is also responsible for compressing, pruning, and quantizing the massive original model to generate a lightweight inference engine suitable for edge deployment, which is then uniformly distributed to all edge nodes.

[0034] Edge nodes: Edge nodes are the front-end units for real-time inference and data acquisition in this invention. They are physically deployed in the DC power distribution network of the smart park, such as in DC gateways, smart combiner boxes, grid-connected inverter controllers of distributed power sources, or edge computing servers. Edge nodes are characterized by their proximity to the data source and extremely low latency. Their core responsibilities are: first, to acquire DC carrier channel data at their location in real time; second, to load a lightweight inference engine distributed by the cloud platform; third, to perform real-time channel modeling inference; and fourth, to perform adaptive adjustment and closed-loop control.

[0035] like Figure 1 As shown, the method disclosed in Embodiment 1 specifically includes the following detailed steps:

[0036] Step 1: Comprehensive Acquisition of DC Carrier Channel Data

[0037] In this embodiment, data acquisition is a comprehensive and targeted process aimed at providing raw data support for subsequent four-path feature extraction. The acquired data is defined as DC carrier channel data, which is a composite dataset that includes at least multi-voltage level basic data and distributed power source data.

[0038] First, regarding the collection of basic data for multiple voltage levels. The DC distribution network of a smart park is typically a complex topology, encompassing different levels of voltage transformation. In this embodiment, we focus specifically on three key voltage levels: 10kV, 380V, and 220V.

[0039] 10kV level: usually corresponds to the medium-voltage DC bus in the park, which is the energy collection point of the main network or microgrid. Its channel characteristics are relatively stable, but it has a large carrying capacity.

[0040] 380V level: This typically corresponds to building-level or regional-level DC power distribution networks, carrying a large number of commercial or industrial loads;

[0041] 220V rating: This typically corresponds to end-point electrical loads, such as DC lighting and office equipment.

[0042] Collecting basic data for these voltage levels includes not only collecting their corresponding electrical parameters such as voltage, current, and power, but more importantly, obtaining information on their network topology, line material, cable length, and node distribution. These are essential for building the physical foundation of the channel model, such as direct transmission paths and load reflection paths.

[0043] Secondly, regarding the collection of distributed generation (DER) data. A core characteristic of smart parks is the high proportion of distributed generation (DER) connections, such as photovoltaic (PV) power sources and energy storage BESS (Battery Safe Energy Storage) power sources. These power sources are connected to the grid via power electronic inverters or converters, and their high-frequency switching behavior generates extremely complex interference and reflections on the DC power line channel. Therefore, this step must collect detailed operational data of these DERs, such as: real-time irradiance and output power of the PV array; real-time charging and discharging power, state of charge (SoC), and impedance characteristics of the grid-connected interface of the energy storage system. The dynamic changes in these data are the main source of channel time-varying characteristics. Data collection can be achieved through sensors deployed at edge nodes, power quality analyzers, smart meters, or by reading directly from the DER controller via a communication bus.

[0044] Step 2: Extraction of four-path channel characteristic parameters

[0045] After collecting massive amounts of raw data, the goal of this step is to analyze this data and transform it from raw time-series signals into structured features that are meaningful to the channel model, i.e., a feature set.

[0046] In this embodiment, the DC carrier channel characteristics of the smart park are mainly determined by the superposition of four key paths. Traditional channel models are severely distorted in DC environments, especially in environments with a large number of DERs. The feature parameters extracted in this embodiment strictly correspond to these four paths:

[0047] 1. Line-of-sight propagation characteristics: This is the line-of-sight propagation path of the signal from the transmitting end to the receiving end. The extracted features mainly include: the physical line length between the transmitting and receiving nodes, the line type, the cable cross-sectional area, and the inherent impedance and attenuation constant determined by the multi-voltage level topology.

[0048] 2. Load Reflection Path Characteristics: This refers to the signal reflection path caused by impedance mismatches resulting from various DC loads connected in the network. Extracted characteristics mainly include: total load power at each voltage level, equivalent impedance of the load type, and changes in load start / stop status.

[0049] 3. Environmental Interference Path Characteristics: This refers to the noise path introduced by external electromagnetic interference (EMI) or line-to-line crosstalk. Extracted characteristics mainly include: the shielding condition of the line, its distance from other high-frequency equipment, and the statistical characteristics of the background noise level.

[0050] 4. Distributed Power Source Reflection Path Characteristics: This is the core interference path that this invention focuses on. It specifically refers to the reflections and noise injection introduced by the grid-connected inverters / converters of DERs such as photovoltaic and energy storage systems. These power electronic devices generate broadband high-frequency switching noise during operation, and their input / output impedances change drastically at different operating points, forming an extremely complex dynamic reflection source. The features extracted in this step are mainly strongly correlated with the DER's operating state, such as the DER's real-time output power, grid-connected interface impedance, inverter switching frequency, and DC bus voltage fluctuations.

[0051] By performing targeted feature engineering on the four paths mentioned above, the system generates a high-dimensional feature set. This feature set is uploaded to a cloud platform for use in the next step of XNN model training.

[0052] like Figure 2 As shown in the diagram, the four-path DC carrier channel model uses DC power lines as the transmission medium (thick black lines in the diagram, labeled with multiple voltage levels of 10kV / 380V / 220V), clearly distinguishing four key signal propagation paths.

[0053] Among them, the direct transmission path (blue solid line) represents the direct line-of-sight propagation path of the signal from the transmitting end to the receiving end. Its continuous solid line shape and strong line width characterize the relatively stable propagation characteristics, corresponding to the inherent parameters of the direct transmission path, such as the physical line length, line type, and cable cross-sectional area between the transmitting and receiving nodes.

[0054] The load reflection path (orange dashed line) shows the signal reflection path caused by impedance mismatch of the load device. The dashed line style vividly represents the discontinuity of the reflected signal and reflects the characteristics of the load reflection path: the signal reflection path caused by impedance mismatch of various DC loads connected in the network.

[0055] The environmental interference path (yellow dotted line) depicts the path through which external electromagnetic interference is introduced. The alternating dotted line pattern appropriately represents the randomness and intermittent nature of environmental interference, corresponding to the characteristics of the environmental interference path: noise paths introduced from external electromagnetic interference (EMI) or line crosstalk.

[0056] The distributed power source reflection path (purple dotted line) vividly illustrates the dynamic reflection characteristics brought about by the high-frequency switching behavior of the DER grid-connected inverter, echoing the characteristics of the distributed power source reflection path: the reflection and noise injection introduced by the grid-connected inverter / converter of DER such as photovoltaic and energy storage, and the core argument that its input / output impedance will change drastically at different operating points.

[0057] Figure 2The color differentiation, line type differences, and line width variations of each path not only enhance visual recognition but also convey the physical differences in signal strength, stability, and time-varying characteristics of each path at a deeper level. In particular, the dynamic fluctuation arrows marked on the DER reflection path confirm the key technical issue mentioned above—that the DER's power fluctuations and interface impedance changes have strong time-varying and random characteristics.

[0058] Step 3: Construction and Application of XNN Channel Model Based on Edge-Cloud Collaboration

[0059] This step is broken down into two stages: cloud training and edge inference, including:

[0060] Phase 1: XNN Model Training on the Cloud Platform. After receiving a large set of features uploaded by the edge nodes as input X, and corresponding measured channel labels, such as measured channel attenuation, signal-to-noise ratio, and bit error rate as output Y, the cloud platform begins training the XNN model.

[0061] XNN Model Structure: First, the interpretable neural network XNN used in this embodiment has a topology intentionally designed to mirror the four-path physical model described above. Specifically, the XNN model is not a single, large, fully connected network, but consists of four parallel subnetworks and a final fusion layer.

[0062] Subnetwork 1: Used to handle features related to the direct transmission path;

[0063] Subnetwork 2: Used to handle features related to the load reflection path;

[0064] Subnetwork 3: Used to handle features related to environmental disturbance paths;

[0065] Subnetwork 4: Used to handle features related to distributed source reflection paths.

[0066] This structural design forms the physical basis for achieving interpretability.

[0067] During training, the cloud platform aims to minimize a specific composite loss function. This loss function aims to achieve sparsity and interpretability of the model while maintaining accuracy. In this embodiment, the specific formula for calculating this loss function is as follows:

[0068]

[0069] in: This character represents the number of samples used for training in this round;

[0070] This character represents the XNN channel model for the first... The channel prediction value output by each training sample. For example, if the measured label is channel attenuation, then... It is the channel attenuation value predicted by the model;

[0071] This character represents the first The actual measured value of the channel corresponding to each training sample, i.e., the label or gold standard;

[0072] This is the first part of the loss function, namely the mean squared error (MSE), which calculates all... The average of the squared differences between the predicted and measured values ​​of each sample;

[0073] This is the second part of the loss function, which is the L1 regularization term for the weights within the subnetwork;

[0074] This character represents the first in the model. Subnetworks ( The internal weight parameters (from 1 to 4, corresponding to the four paths respectively);

[0075] This indicates a subnetwork All weight parameters Calculate its L1 norm, which is the sum of the absolute values ​​of all weight parameters;

[0076] This character is a hyperparameter, namely the first regularization penalty coefficient. It is used to balance the relationship between accuracy (MSE term) and subnetwork sparsity. By penalizing non-zero weights, this term forces the model to automatically prune feature connections within the four subnetworks that do not contribute much to the results during training, making each subnetwork more concise and sparse, which helps with feature selection and prevents overfitting;

[0077] This is the third part of the loss function, which is the L1 regularization term for the path fusion coefficients;

[0078] This character represents a vector, specifically the path correlation coefficient. If the final fusion layer of the XNN is a weighted summation layer, then... It refers to the fusion weights corresponding to the outputs of the four sub-networks, which represent the contribution or importance of each of the four paths to the final channel state.

[0079] This indicates the path correlation coefficient. For example, calculate the L1 norm of a vector containing 4 elements;

[0080] This character represents the second regularization penalty coefficient, used to balance accuracy and path sparsity. This term is the most critical step in achieving interpretability, through... Applying L1 penalty forces the model to prioritize the correlation coefficients of less important paths during training. Compressed to very close to zero. For example, under certain operating conditions, if environmental interference, such as path 3, is very small, then its corresponding... It will then approach 0. Ultimately, the model will automatically learn which paths are the dominant paths in a specific scenario.

[0081] Once the cloud platform trains the XNN model to convergence using the aforementioned loss function, it obtains a large but sparse original model. Subsequently, the cloud platform performs model compression operations, such as weight pruning and parameter quantization, to convert the model into a lightweight inference engine with a small size and low computational cost, and then distributes it to edge nodes deployed throughout the campus via the network.

[0082] Phase Two: XNN Model Inference at Edge Nodes. After receiving the lightweight inference engine from the cloud, the edge nodes load it into their local memory and begin real-time channel modeling inference. This step details how the edge performs efficient inference while maintaining interpretability.

[0083] First, the core feature of this loaded lightweight inference engine is that it fully preserves the core structure of the four sub-networks in the XNN channel model, corresponding to the direct transmission path, load reflection path, environmental interference path, and distributed power source reflection path, respectively. This means that the compression process does not compromise the fundamental interpretability of the model. The model at the edge nodes remains a four-branch parallel structure.

[0084] Secondly, the specific process of inference performed by edge nodes is as follows:

[0085] (1) Parallel Input: The edge nodes input their local feature parameters, which are collected and extracted in real time, into the four sub-networks in parallel and non-blocking manner according to their physical meaning. For example, the line length feature is sent to sub-network 1, while the power fluctuation parameter collected in real time is sent to sub-network 2. The features are then fed into subnetwork 4;

[0086] (2) Obtain the influence components separately: Since the four sub-networks are structurally separated, they will be calculated independently to obtain the channel influence components of the corresponding four paths.

[0087] For example, in At any given moment, the output of subnetwork 1 might be -10dB, representing direct transmission attenuation; the output of subnetwork 2 might be -3dB, representing load reflection attenuation; the output of subnetwork 3 might be -1dB, representing environmental noise interference; and the output of subnetwork 4 might be -15dB, representing strong reflection interference from the DER.

[0088] (3) Finally, the inference engine performs weighted fusion of the four channel influence components, such as using the path correlation coefficients trained in the cloud. The weighted summation is performed to obtain the final channel prediction result, such as a total attenuation of -29dB.

[0089] While the edge nodes arrive at the final result of -29dB, they also clearly retain the four components that led to this result: -10, -3, -1, and -15.

[0090] This enables interpretable, real-time reasoning about the channel state at the edge. When the system administrator or AI sees this result, it not only knows that the channel is poor (-29dB), but also knows more precisely that the main cause of the poor channel is the distributed power source reflection path (-15dB).

[0091] like Figure 3 The real-time output component waterfall plot of the four sub-networks of XNN is shown, which intuitively presents the dynamic changes of the attenuation components of the four sub-networks (corresponding to the four paths) in the multi-voltage level DC carrier channel of the smart park within a 300-second monitoring period. Figure 3 The graph uses time t (0~300s) as the horizontal axis, path number (1~4) as the vertical axis, and attenuation component (unit: dB) as the Z-axis. It uses a Parula color map to gradually change from dark blue (low attenuation) to yellow (high attenuation). Figure 3 The independent output components of the four paths are separated and displayed through a three-dimensional perspective and color gradient.

[0092] Specifically, the output of path 1 (direct transmission main path) is stable at around -10dB, with only ±0.5dB random fluctuations, reflecting the low-loss characteristics of the DC bus near-field coupling; path 2 (load reflection path) exhibits a -12dB reference value superimposed with ±2dB periodic oscillations (frequency 0.01Hz), accurately corresponding to the characteristics of high-frequency reflection oscillations caused by load abrupt changes; the output of path 3 (environmental interference path) is a -15dB reference value superimposed with ±1.5dB Gaussian noise, reflecting the statistical law of random disturbances in the electromagnetic environment of the park; path 4 (DER reflection path) is stable at -8dB when t<180s, but suddenly rises to -2dB after t=180s (with ±0.2dB fluctuations), directly confirming the dominant contribution of impedance mismatch in the distributed energy inverter DER to channel degradation.

[0093] Step 4: Monitor fluctuations and make dynamic adjustments

[0094] This method also includes a step of monitoring the fluctuations in distributed power source data. When the fluctuations meet the preset adjustment conditions, the parameters of the XNN channel model are dynamically adjusted to achieve adaptive updates of the channel model.

[0095] In summary, this first embodiment comprehensively constructs the core foundation of a multi-voltage-level DC carrier adaptive channel modeling method for smart parks. By utilizing an edge-cloud collaborative architecture, it addresses the computational power requirements for model training and the real-time requirements for edge inference. Furthermore, through the innovative design of a four-path mirrored XNN structure and a composite L1 regularized loss function, it trains a physically interpretable model in the cloud. Finally, by retaining a lightweight engine for parallel inference across four sub-networks at the edge, it successfully achieves interpretable real-time channel diagnostics on resource-constrained edge nodes.

[0096] Example 2:

[0097] The core objective of this embodiment is to elaborate and define in detail how the aforementioned dynamic adjustment steps are fully implemented in a highly reliable system. The intelligent adaptive mechanism disclosed in this embodiment is crucial for ensuring the long-term accuracy of the XNN model in Embodiment 1 within the complex and ever-changing DC environment of a smart park. Without this mechanism, the model trained and deployed from the cloud would quickly become outdated as the operating conditions of the distributed power supply (DER) shift, leading to a simultaneous decline in the model's prediction accuracy and interpretability. The complete technical solution of this embodiment, from the perception layer and decision layer to the execution layer, defines in detail how edge nodes autonomously achieve model self-evolution and dynamic adaptation.

[0098] Part 1: Adaptive Perception Layer

[0099] In this invention, the primary cause of drift in the channel model, especially in the parameters of the distributed generation reflection path network, is the change in the operating state of the DER (photovoltaic, energy storage). Therefore, this embodiment first defines two core physical input parameters for driving the adaptive mechanism: distributed generation data.

[0100] I. Definition and Acquisition of Core Physical Parameters

[0101] Edge nodes are configured to monitor and calculate the following two key parameters in real time at a high frequency, such as a sampling rate of 10 to 1000 times per second:

[0102] 1. Power fluctuation parameters : This is the difference between the real-time grid-connected power and the rated power of the distributed power source. The real-time grid-connected power is an instantaneous value obtained by directly reading the internal power metering register of the edge node through the real-time communication bus between the edge node and the DER controller. The rated power, on the other hand, is a static configuration parameter, such as... It is written into the configuration file when the edge node is deployed, representing the full load capacity of the DER design.

[0103] This parameter is the most direct indicator of the shift in the operating point of the DER. The input and output impedances and switching harmonic spectrum of the DER, especially its core power electronic converter, are strongly correlated with its output power, i.e., its operating point. Therefore, Dramatic changes in the channel characteristics are strong precursor signals that a fundamental shift in channel properties is imminent.

[0104] 2. Interface impedance parameters : This is the difference between the real-time impedance of the grid connection interface of the distributed power source and the line reference impedance. Line reference impedance. This is a complex number representing the equivalent impedance of a DC power line at the grid connection point (PCC) under standard operating conditions. This value is typically obtained through simulation calculations during the park design phase, or through practical calibration using specialized instruments during the system commissioning phase, and is then stored as a reference value in the non-volatile memory of the edge node.

[0105] If If it's a phenomenon, then... That's the essence. Impedance mismatch is the physical root cause of signal reflection. The drastic changes directly mean that the reflection coefficient and phase of the distributed power source reflection path have changed, which has a fatal and direct impact on the channel model. Simultaneously monitoring... and This constitutes a double verification of the DER state.

[0106] II. Engineering from Physical Parameters to Model Features

[0107] Raw data collected from edge nodes and These parameters cannot be directly used by the XNN subnetwork in Example 1. The model requires normalized features with clear physical meaning. Therefore, this example further defines how the feature engineering module inside the edge nodes calculates these raw parameters into features usable by the XNN. These features, namely the feature parameters used to characterize the distributed power source reflection path, are adaptively adjusted inputs and also the inputs for the XNN model inference in Example 1.

[0108] 1. Reflection power factor This feature is used to quantify the impact of power fluctuations on reflection intensity, and its calculation formula can be expressed as:

[0109]

[0110] in: : That is, the reflection power factor feature value calculated in real time by the edge node, which will be sent into the XNN sub-network 4;

[0111] This refers to the base reflection power factor. This is a baseline constant, also calibrated and configured during system deployment. It represents the performance of the DER when it is in a standard state, such as... The reflection level of that path at that time;

[0112] That is, to The absolute value is used. This reflects an important technical understanding: whether the power increases or decreases, it results in a drastic shift in the operating point, which will lead to a change in impedance. Therefore, the absolute value is used to characterize the magnitude of the shift.

[0113] : That is, the rated power, which is used as the normalization denominator here, so that It becomes a dimensionless power offset rate in the range [0,1].

[0114] This is a power-reflection sensitivity coefficient. This coefficient is not set out of thin air; it is an empirical value obtained through regression analysis and model optimization of massive historical data during the cloud training phase of Example 1. It is embedded in the model parameter package sent to the edge nodes and represents the average impact weight of power changes of this specific DER model on channel reflection.

[0115] 2. Impedance correction factor This feature is used to quantify impedance mismatch correction in the channel model, and its calculation formula can be expressed as:

[0116]

[0117] in: : That is, the impedance correction coefficient characteristic value calculated in real time by the edge node, which will also be sent to the XNN sub-network 4;

[0118] This is a reference value here, representing the condition when the impedance is unbiased, i.e. When the correction factor is 1, no correction is performed;

[0119] That is, to The absolute value represents the magnitude of the impedance deviation.

[0120] The line reference impedance is used as the normalization denominator here.

[0121] This is an impedance-reflection sensitivity coefficient, and... Similarly, it is also a product of cloud-based model training, representing the average impact weight of normalized impedance deviation on channel reflection.

[0122] Through the above steps, edge nodes will handle high-frequency changes in the physical world. , It was transformed into low-frequency, normalized, physically meaningful characteristics. , The construction of the perception layer has been completed.

[0123] like Figure 4 The normalized trajectory diagram of the distributed source reflection path is shown, clearly revealing how the DER reflection path changes under the influence of the reflection power factor in the high-frequency physical world. (Horizontal axis component) and impedance correction factor The dynamic evolution law of the characteristic phase trajectory composed of (vertical axis components) transforms complex nonlinear coupling into a low-dimensional interpretable two-dimensional normalized space. Figure 4 The medium-to-dark orange trajectory extends outward from the origin along a nonlinear path, reflecting the overall interference intensity. The trajectory exhibits a spiral decay pattern within the r=1 unit circle, precisely corresponding to the physical mechanism of the interaction between high-frequency impedance mismatch and power fluctuations. Green dots mark a stable state, indicating that the DER is grid-connected and synchronized without reflection. Yellow squares correspond to moderate disturbances, reflecting the characteristic drift under light load switching or small signal disturbances. Red triangles indicate strong reflection mismatch, directly confirming the extreme operating condition where sudden changes in inverter power lead to impedance collapse and a surge in reflections.

[0124] Figure 4 The trajectory remained within the unit circle and tended to saturate, verifying the normalization boundedness and controllability of feature engineering, and providing a quantitative visualization basis for subsequent composite statistical triggering conditions.

[0125] Part Two: Adaptive Decision-Making Layer

[0126] Having completed the perception process, the next question is when to trigger the adaptive adjustment. However, if a poorly designed trigger is used, such as simply... Even the slightest change triggering an adjustment can lead to system disaster. It can cause the model parameters of edge nodes, especially the weights of subnetworks, to oscillate at high frequencies. This model jitter is worse than model drift itself because it prevents communication protocols from establishing stable links.

[0127] Therefore, the triggering mechanism defined in this embodiment is a robust, anti-jitter composite statistical triggering condition designed to identify real state transitions rather than instantaneous noise. This decision logic runs continuously as a background service on the edge node.

[0128] I. Composite Statistical Triggering Conditions

[0129] This condition is an OR logic block; the decision layer will output a trigger signal if any of the following conditions are met:

[0130] Condition A: Moving average statistics based on power fluctuations

[0131] Definition: In the first preset time window Internal power fluctuation parameters The moving average is greater than the first power average threshold.

[0132] It is important to note that edge nodes maintain a [database name] in memory. A sliding time window, like a FIFO (First In First Out) queue;

[0133] That is, the first preset time window, for example Seconds, meaning the queue always retains the data from the past 5 minutes. Sampled values;

[0134] Moving average: The edge node calculates the average value of all elements in the queue every second. The sample can be calculated using either the exponential moving average (EMA) or the simple moving average (SMA). The EMA is considered a better option because it has a higher weighting on new data.

[0135] First average power threshold: This is a key configuration threshold, such as setting it to... ;

[0136] Triggering logic: such as when the past 5 minutes When the EMA value consistently exceeds 30% of the rated power, the trigger determines that this is not a transient disturbance, but rather a fundamental and persistent change in the DER's operating mode. In this case, adaptive activation must be triggered.

[0137] Condition B: Statistical analysis of continuous hold time based on impedance deviation

[0138] Definition: Second preset time window Internal impedance correction factor The duration for which the impedance remains above the first impedance correction threshold is greater than a preset duration threshold. .

[0139] It should be noted that this is a state persistence detector, used to capture operating conditions where power fluctuations are small, but impedance characteristics have undergone a fixed shift.

[0140] That is, the second preset time window, for example Seconds, this is the total observation window;

[0141] First impedance correction threshold: A threshold used to determine mismatch, for example, set to 1.15;

[0142] That is, a preset duration threshold, for example Second;

[0143] Triggering logic: The edge node starts a persistent state timer, and when it detects... When the timer starts, the countdown begins.

[0144] If the next second still The timer continues; if there is one second in between... The timer is immediately reset to zero;

[0145] When the accumulated time of the timer (which has not been reset) exceeds At 60 seconds (for example), the trigger determines:

[0146] The grid-connected impedance of the DER has stably shifted to a new, erroneous state, and this state has persisted for at least one minute. This is also a strong signal that the model must be updated. Through this combined statistical decision of moving average and duration, this embodiment ensures that adaptive adjustments are only activated when truly needed, greatly guaranteeing system stability.

[0147] Part Three: Adaptive Execution Layer

[0148] Once the decision-making layer issues a trigger signal, the execution layer begins its work, with the goal of modifying the XNN model parameters in Implementation Example 1.

[0149] I. Calculation of Nonlinear Adaptive Adjustment Factor

[0150] The first step in adjustment is to determine how much to adjust, i.e., the magnitude of the adjustment. This embodiment introduces a core variable, an adaptive adjustment factor. :according to moving average or Calculate a nonlinear adaptive adjustment factor based on the duration exceeding the threshold. ,in It is positively correlated with the moving average or the duration of exceeding the threshold. The calculations are non-linear to enable fine-tuning for small deviations and readjustment for large deviations, but with an upper limit on the control logic. Consider the following scenario:

[0151] Scenario A ( Triggered): If triggered by condition A, The calculation can be performed using an S-shaped function, and the formula can be expressed as:

[0152]

[0153] in, This is the maximum adjustment range, such as 0.5; It is the gain coefficient; This is the power average threshold. This function ensures that when... When it exceeds the threshold, It will increase rapidly, but will eventually saturate. to prevent The infinite increase of ...

[0154] Scene B ( Triggered): If triggered by condition B, The calculation can be positively correlated with the duration of exceeding the threshold, and the formula can be expressed as:

[0155]

[0156] in, It is the basic adjustment amount; It is gain; This is the actual duration. The function, which always uses a logarithmic function, also achieves a positive correlation but with slower, non-linear growth.

[0157] like Figure 5 It demonstrates the nonlinear adaptive adjustment factor The S-curve plot depicts the adaptive adjustment factor. Nonlinear response mechanism to normalization bias. Figure 5 The four S-curves (k=6, 10, 14, 20) are represented by a gradient from dark blue to bright green to indicate increasing sensitivity. The black dashed line marks the trigger threshold of 0.3, i.e., 30% of the rated power. The yellow pentagram at (0.3, 0.5) is the inflection point of the S-curve, and the purple square at (0.6, 0.5) is for illustration. Saturation median; when x < 0.3 ≈0, zero adjustment at this point to avoid accidental activation; when x=0.3 The value jumps to 0.5, triggering an instantaneous amplification; after x > 0.6... When the value approaches 1, it becomes saturated and locked to prevent over-adjustment.

[0158] II. Target Weight Amplification

[0159] Calculate the adjustment range Then, the execution layer begins to modify the model parameters. Crucially, the XNN model of this invention is interpretable, i.e., a four-path subnetwork. Therefore, the adjustments must be targeted.

[0160] Specifically, the adaptive adjustment factor is applied. This amplifies the output weights of the sub-networks corresponding to the distributed source reflection paths in the XNN channel model. In actual implementation, recalling Example 1, the inference of edge nodes involves a weighted fusion of the outputs of the four sub-networks. These weights are the path correlation coefficients trained in Example 1. Vectors, for example These correspond to direct transmission, load balancing, environment, and DER, respectively.

[0161] Because the root cause of this trigger is DER ( or Therefore, the execution layer only adjusts the weights of the corresponding DER. The formula is adjusted as follows:

[0162]

[0163] Among them, here This is the weight value currently being used by the edge node; It is a positive number, therefore the operation amplifies it. .

[0164] The significance of the above formula lies in the fact that it is physically reasonable. The trigger detects that the fluctuation of DER has intensified, which means that the impact of the distributed power source reflection path on the total channel has increased. Therefore, the model must increase the weight of this path in the final fusion result, which is the great advantage brought by the interpretable model.

[0165] III. Stability Control

[0166] If only the previous step is executed, i.e., amplification... This will have a serious side effect: the total energy of the model's output will change. Let's assume the original... ,enlarge After, the sum This leads to a systematic inflation of the model's predictions for all inputs, compromising the model's stability. To address this issue, normalization reduction needs to be applied to the output weights of the other three sub-networks in the XNN channel model to maintain the stability of the model's output. Specifically:

[0167] Objective: To increase At the same time, maintain Constant;

[0168] Step 1: Calculation Increment ;

[0169] Step 2: This increment It must be deducted from the weights of the other three sub-networks;

[0170] Step 3: Deduct proportionally and calculate the old percentages for the other three weights: ;

[0171] Step 4: Perform normalization reduction: ;

[0172] Through this operation ;

[0173] And the new sum .

[0174] In this way, without changing the total output energy of the model, the edge nodes have completed the precise transfer of weights from the direct transmission, load, and environment paths to the DER path, perfectly achieving targeted adaptation.

[0175] In summary, the intelligent adaptive mechanism disclosed in Embodiment 2 solves the problem of what to perceive through defined physical parameters and feature engineering; it solves the stability problem of when to adjust through composite statistical triggering; and finally, it solves the accuracy problem of how to adjust through nonlinear factors, targeted amplification, and normalized reduction. This enables the XNN model of Embodiment 1 to evolve into a dynamic adaptive system capable of tracking the dynamics of the park in real time.

[0176] Example 3:

[0177] The preventive closed-loop control method disclosed in this embodiment has a complete and implementable execution process that includes the following four closely linked stages:

[0178] Phase 1: Proactive Prediction of Communication Risks

[0179] 1. Acquisition of Predictive Data Sources. To predict the future, edge nodes must acquire future input features. In the specific scenario of a smart park, this is highly feasible. In this embodiment, the edge node is configured to actively access and subscribe to the future scheduling database of the park's Energy Management System (EMS) or Building Automation System (BAS) via its uplink communication interface. Specifically, the edge node will acquire at least the following two key predictive inputs in real time:

[0180] DER scheduling plan: For example, the BESS charge and discharge power curve of the energy storage system and the ultra-short-term forecast curve of photovoltaic power generation issued by the EMS control room for the next hour;

[0181] High-power load start-stop plans: such as load scheduling events preset in the EMS control room or BEMS, such as starting the central air conditioning unit at 2 pm and turning on all DC fast charging piles in the park at 3 pm.

[0182] 2. Prediction Time Window The edge nodes will have a forward-looking time window set, i.e. , The settings have the following limitations:

[0183] if If the warning time is too short, such as 1 minute, the warning time is insufficient, and physical systems with large loads may not be able to respond to control commands in time.

[0184] if If the timeframe is too long, such as 2 hours, the uncertainty of the EMS forecast data itself increases, leading to false warnings.

[0185] In this embodiment, a method with high implementation value is described. It was set to 15 minutes.

[0186] 3. Rolling Prediction and Risk Assessment. Edge nodes execute rolling prediction tasks on their local processors.

[0187] For example, at the current time 14:00:00, the edge node obtains all EMS scheduling plans between 14:00:00 and 14:15:00.

[0188] It will change these future DER power variations and load start / stop events As input features, these features are fed into the XNN model, which has been fully calibrated locally by Example 2. The XNN model then outputs the predicted channel state values ​​for each second (or each minute) over the next 15 minutes, such as the predicted signal-to-noise ratio. .

[0189] 4. Assessment of degradation risk. Edge nodes will With a pre-configured QoS alarm threshold (such as...) The comparison is performed using a value of 15dB to determine if there is a risk of degradation. The specific engineering definition of a risk of degradation is: in the future... Within the time window, there exists at least one Time makes Once this condition is triggered, for example, if the XNN model predicts that the signal-to-noise ratio will drop to 12dB at 14:10:30, the system immediately determines that there is a risk of degradation and immediately switches to the next stage.

[0190] Phase Two: Explainable Diagnosis of the Root Causes of Deterioration

[0191] When the risk of degradation in the first stage is confirmed When a time event occurs, the edge nodes do not only consider the XNN model. The final fused output at time step. Instead, it retrieves the intermediate layer output from within the XNN model, specifically the predicted values ​​of the four channel influence components before fusion in Example 1:

[0192] 1. Diameter influence component (predicted value): ;

[0193] 2. Load reflection path influence component (predicted value): ;

[0194] 3. Environmental disturbance path influence component (predicted value): ;

[0195] 4. Influence component of distributed generation reflection path (predicted value): ;

[0196] The system will compare the deterioration contribution of these four components. For example, by calculating which component... The system determines the moment of greatest degradation, or the absolute degradation value of the component that is greatest. Then, it arrives at a unique and actionable diagnostic conclusion:

[0197] Conclusion A: This is the dominant interference path, which means that after 10 minutes, the power slope of the energy storage BESS is too large, causing the DER reflection path to collapse.

[0198] Conclusion B: This is the dominant interference path, which means that the EV charging piles will start up simultaneously 10 minutes later, causing the load reflection path to collapse.

[0199] This diagnostic result is the sole basis for subsequent precise control.

[0200] Phase 3: Physical Layer Control (A): Power Smoothing for DER

[0201] When the diagnosis conclusion of the second stage is Conclusion A, that is, when the dominant interference path is the distributed power source reflection path, a power smoothing command is sent to the controller of the distributed power source. Specifically:

[0202] The edge node addresses the controller of the distributed power source causing the problem directly through its control bus; the power smoothing command sent is a fine-grained slope control command that includes: the original plan, the diagnostic results of the XNN, the calculations of the edge node, and the specific instruction content.

[0203] Upon receiving this instruction, BESS's PCS controller treats it as a local constraint, overriding the EMS's global instruction. As a result, due to the physical world... Dramatic changes are smoothed out, and actual interference on the DC channel is suppressed above the communication QoS threshold, thus proactively avoiding potential communication interruptions.

[0204] Phase 4: Physical Layer Control (B): Load Shifting of EMS

[0205] When the diagnosis conclusion of the second stage is Conclusion B, that is, when the dominant interference path is the load reflection path, a load time-shift scheduling command is sent to the park's energy management system. Specifically:

[0206] In this scenario, the edge node does not directly control that load. Instead, it sends a scheduling coordination request to its superior, the campus's Energy Management System (EMS). This request, rather than a command, specifies the time and reason for its action.

[0207] The EMS central scheduler received this request:

[0208] Scenario 1 (EMS Accepts): If EMS determines that it is not a rigid requirement, EMS will accept the request and postpone the execution of the task.

[0209] Scenario 2 (EMS optimization): EMS determines that the task cannot be delayed, but it can be broken down into smaller parts, i.e., executed in batches.

[0210] Regardless of whether EMS adopts scenario 1 or scenario 2, the instantaneous, massive load reflection impact that originally occurred at 14:10:30 is physically shifted or distributed. Therefore, the predicted channel degradation is actively avoided.

[0211] like Figure 6 The graph shows a rolling prediction of signal-to-noise ratio (SNR) with a prediction time window of 15 minutes. It demonstrates the continuous rolling prediction trend of SNR within the next 15 minutes by the XNN model based on the current channel state. The horizontal axis represents the look-ahead time, the vertical axis represents the predicted SNR, and the black dashed line marks the service quality threshold of 15 dB.

[0212] The prediction curve starts from around 18 dB, with slight periodic fluctuations, and slowly decreases to around 16 dB in the first 10 minutes, which is in line with the expectation of stable transmission under the normal load change of the park. However, from the 10.5-minute mark, the signal-to-noise ratio drops rapidly, and by the 12-minute mark, it has dropped below 12 dB. The red transparent filler area accurately covers this degradation warning range, which directly corresponds to the prospective diagnosis example of the predicted signal-to-noise ratio of 12 dB to less than 15 dB mentioned above at 14:10:30. The red filler block further highlights the warning starting point, indicating that the system needs to immediately start feedforward adjustment.

[0213] Figure 6 By transforming the abstract rolling prediction and proactive early warning mechanism into intuitive time-series evidence, this invention demonstrates that it can detect risks 2-3 minutes before actual channel degradation, providing timely and reliable decision-making basis for subsequent feedforward optimization of three communication parameters.

[0214] In summary, this third embodiment upgrades the XNN model from a passive channel analyzer to a proactive communication risk controller through a complete process of prediction-diagnosis-precise control. This solution no longer waits for communication interruptions to retransmit data, but proactively eliminates risks by controlling physical interference sources before interruptions occur. This demonstrates an unprecedented depth of collaboration between information and physical systems, significantly improving the proactive reliability and predictability of critical communication links in smart parks.

[0215] Example 4:

[0216] This embodiment, within the multi-voltage-level DC carrier adaptive channel modeling method for smart parks, serves as the final intelligent defense to ensure the robustness of the communication link in scenarios where physical layer control command execution fails. The core objective of this embodiment is to demonstrate how, in the preventative closed-loop control described in Embodiment 3, when the power smoothing command issued to the distributed power source fails to execute successfully for various reasons, the system can mitigate impending communication quality degradation through proactive parameter adjustment at the communication layer, thereby achieving maximum availability of the communication system in complex power environments.

[0217] In the actual operation of a smart park, the control module of a distributed power source may refuse or be unable to execute power smoothing commands issued by edge nodes for various reasons. For example, the distributed power source may be in a forced output mode with higher grid dispatch priority, and its local controller may refuse the smoothing command based on safety constraints; or, the communication link between the control module and the edge node may experience a momentary interruption, resulting in the command failing to be delivered or the acknowledgment signal being lost; or, although the control module receives the command, it may not actually execute it due to internal logic errors. Regardless of the reason, if the system relies solely on physical layer control, in such scenarios of command execution failure, the predicted channel degradation will inevitably occur, leading to communication interruption. Therefore, the communication layer adaptive avoidance strategy in this embodiment, as a supplement and backup to physical layer control, is crucial.

[0218] The complete technical solution disclosed in this embodiment four includes the following key stages: continuous monitoring and confirmation of instruction execution status, intelligent judgment of instruction execution failure, and precise execution of the communication layer adaptive avoidance strategy.

[0219] Phase 1: Continuous monitoring and confirmation of instruction execution status

[0220] When the edge node identifies the distributed power source reflection path as the dominant path for channel degradation in the future based on the preventive closed-loop control process of Embodiment 3, and issues a power smoothing command to the specific distributed power source control module accordingly, the monitoring and confirmation mechanism in this stage is immediately activated.

[0221] When an edge node issues a power smoothing command, it simultaneously creates a command execution acknowledgment timer in its local memory. This timer is a countdown timer with a configurable timeout parameter, typically set based on the response characteristics of the distributed power source and communication latency. For example, in a typical smart campus scenario, the timeout can be set to a value between 5 and 30 seconds to ensure sufficient time to wait for a response without delaying the avoidance mechanism due to excessive waiting.

[0222] During the timer's execution, the edge node will perform two monitoring tasks in parallel:

[0223] The first task is to monitor its communication interface and wait for a response from the distributed power control module. There are two possible response types: one is a command execution confirmation response, indicating that the control module has received and committed to executing the power smoothing command; the other is a command rejection response, indicating that the control module is unable to execute the command for some reason.

[0224] The second task is to continue monitoring the power fluctuation parameter ΔP of the distributed power source in real time. Edge nodes will continuously acquire the real-time grid-connected power of the distributed power source at a high sampling rate and calculate the difference between it and the rated power. Observe at the same time Does the trend of change converge in the direction expected by the power smoothing command? For example, if the power smoothing command requires reducing the slope of the power change, then the edge nodes would expect to detect... The fluctuation range gradually decreases or the rate of change tends to level off.

[0225] Phase Two: Intelligent Judgment of Instruction Execution Failure

[0226] When the instruction execution acknowledgment timer times out, or when a specific response is received before its timeout, the edge node initiates intelligent judgment logic to ultimately determine whether the power smoothing instruction has failed. This judgment logic is based on either of the following two conditions:

[0227] Condition 1: Before the instruction execution confirmation timer expires, the edge node receives an instruction rejection response from the distributed power control module. Upon receiving such an explicit rejection response, the edge node immediately determines that the power smoothing instruction execution has failed. This is a fast-failure mechanism that allows the system to initiate avoidance strategies as early as possible without waiting for the timer to expire.

[0228] Condition 2: The command execution confirmation timer times out, and the edge node has not received any form of execution confirmation response by the timeout point. In this case, the edge node will not immediately determine failure, but will initiate a supplementary verification process. This process requires the edge node to check the power fluctuation parameters monitored throughout the entire period from when the command was issued to the present. Whether it converges in the expected direction of the power smoothing command.

[0229] Specifically, the edge node invokes an internal convergence determination algorithm. This algorithm can calculate... The moving average, standard deviation, or slope of change within the time window after the instruction is issued is compared with the baseline value within the same time window before the instruction was issued, or with a preset convergence threshold. If the comparison result indicates... If the fluctuations do not converge as expected, for example, if the fluctuation amplitude does not decrease or even increases, the edge node determines that the power smoothing instruction has failed to execute.

[0230] Only when either condition one or condition two is met does the system formally determine that the physical layer's power smoothing control has failed, and immediately trigger the third-stage adaptive avoidance strategy at the communication layer. This dual-judgment mechanism effectively avoids misjudgments caused by communication delays and also prevents [further issues]. Even when spontaneous improvements occur by chance, avoidance strategies are not activated unnecessarily, thus ensuring the accuracy of system decisions.

[0231] Phase 3: Precise Execution of Adaptive Avoidance Strategies at the Communication Layer

[0232] Once the instruction execution failure is confirmed, the edge node will immediately cease waiting for physical layer control and instead initiate a communication layer adaptive avoidance strategy. The core idea of ​​this strategy is to acknowledge the reality that physical layer interference cannot be suppressed, and instead proactively adapt to the upcoming high-interference environment by adjusting the parameters of the communication system itself in advance, thereby maintaining the availability of the communication link. The entire strategy execution consists of two core steps:

[0233] Step 1: Lock onto the high-interference prediction state

[0234] The edge node first performs a state labeling operation on its internally held interpretable neural network XNN channel model. Specifically, it locks the output state of the sub-network corresponding to the distributed source reflection path in the XNN channel model. In the normal inference process, the output of this sub-network is a continuously changing channel influence component value. However, in this step, the edge node forcibly labels the output of this sub-network as a fixed high-interference prediction state. This is not a modification of the model's own computational logic, but a post-processing annotation of the model's output, designed to explicitly indicate to the subsequent communication decision module that the distributed source reflection path will be regarded as a persistent and high-intensity interference source in the present and for some time to come.

[0235] This high-interference prediction state is a qualitative state indicator, but it can be associated with one or more quantitative levels of interference.

[0236] For example, during system initialization, a signal-to-noise ratio (SNR) degradation threshold can be defined through historical data analysis or simulation. When the predicted SNR falls below this threshold, a high-interference state is triggered. When the edge node locks into this state, it can classify the distributed power source reflection path into a preset high-interference level based on the most recent predicted output value of the XNN model.

[0237] Step 2: Issue communication parameter pre-adjustment instructions

[0238] Based on the aforementioned high-interference prediction state, edge nodes no longer passively wait for actual communication quality degradation before taking action. Instead, they proactively and feedforward send communication parameter pre-adjustment commands to the physical transceiver carrying DC carrier communication. These commands aim to achieve higher robustness of the communication link by reducing communication rate, improving error correction capabilities, or avoiding interference bands. The pre-adjustment commands include one or more combinations of the following:

[0239] The first adjustment command is to reduce the modulation order of the modulation and demodulation. DC carrier communication typically uses multi-order modulation schemes, such as QPSK, 16QAM, and 64QAM. The higher the modulation order, the higher the spectral efficiency, but the more stringent the requirements for the channel signal-to-noise ratio. Under high interference prediction conditions, the edge node sends a command to the transceiver, requesting it to reduce the modulation order by one or more levels.

[0240] For example, downgrading from 64QAM to 16QAM, or from 16QAM to QPSK. While this downgrading reduces the peak data transmission rate, it significantly improves demodulation success rate under harsh channel conditions, thus avoiding communication interruptions. An internal mapping table of modulation order and interference level can be stored within the edge node to select an appropriate modulation scheme based on the severity of the interference.

[0241] The second adjustment directive is to increase the coding redundancy level of Forward Error Correction (FEC). Forward Error Correction (FEC) technology adds redundant check bits to the data stream, enabling the receiver to automatically correct errors within a certain error range. A higher coding redundancy level provides stronger error correction capability, but correspondingly reduces effective data throughput. Under high interference prediction conditions, edge nodes instruct transceivers to increase the redundancy level of their FEC coding. This can be achieved by switching from a coding scheme with higher efficiency to one with lower efficiency but stronger error correction capability, such as switching from Lisso codes to a concatenated code of convolutional and Lisso codes, or by directly increasing the length of the check bits. This measure further enhances the communication system's fault tolerance against sudden and continuous interference.

[0242] The third adjustment command is to switch the communication subcarrier to the backup frequency band that has the least impact under high interference prediction conditions. DC carrier communication typically operates within a certain frequency band and may employ multi-carrier technology. Different interference sources have varying degrees of impact on different frequency bands. Based on its historical learning data from its XNN model or a pre-stored frequency band-interference feature database, the edge node can identify which backup frequency band is least affected by distributed power source reflection paths under the current high interference prediction conditions. Subsequently, it issues a command to the transceiver, requesting that the primary communication subcarrier or the entire communication frequency band be switched to this backup frequency band.

[0243] For example, if diagnostics reveal that interference is primarily concentrated in the low-frequency band, the system instructs a switch to the high-frequency band for communication. This band avoidance strategy effectively avoids the areas with the most severe interference.

[0244] The issuance of the aforementioned communication parameter pre-adjustment command is an atomic operation. The edge node sends the command to the transceiver in the form of a high-priority command frame through its underlying communication driver interface. After receiving the command, the transceiver will complete the parameter reconfiguration within one of its internal communication parameter update cycles, thereby switching to the new communication mode almost seamlessly.

[0245] To ensure the effectiveness of the mitigation strategy, the edge node does not completely ignore the issue after issuing the pre-adjustment command. It continues to monitor actual channel quality metrics, such as real-time bit error rate or signal-to-noise ratio. If it finds that communication quality continues to deteriorate even after parameter adjustments, the edge node can initiate further mitigation measures, such as combining multiple adjustment commands mentioned above, or triggering system-level alarms to notify operations and maintenance personnel to intervene.

[0246] like Figure 7 The simulation comparison chart of BER under three strategies under high interference conditions is shown. The performance improvement of bit error rate by the three-way feedforward adjustment of communication parameters is intuitively compared in a high interference campus environment. The horizontal axis is the signal-to-noise ratio and the vertical axis is the bit error rate.

[0247] The red dashed line represents the original high-order modulation method, with a bit error rate as high as one percent at a signal-to-noise ratio of 10 dB; the blue dotted line represents the order reduction combined with a strong forward error correction strategy, which drastically reduces the bit error rate to one millionth at the same signal-to-noise ratio; the green dotted line further superimposes frequency band switching to avoid distributed energy harmonics, and under the same conditions, the bit error rate is as low as one ten-millionth. The triangle markings accurately correspond to the above key performance points.

[0248] Figure 7The hierarchical gain effect of the three strategies described in Example 4—modulation order reduction, forward error correction enhancement, and intelligent frequency band switching—is clearly demonstrated. This enables the system to quickly recover from an unusable state to high-reliability transmission when the channel deteriorates, verifying the fast response and significant robustness of the feedforward adjustment. It transforms the abstract parameter optimization combination into a reproducible engineering performance curve.

[0249] The adaptive avoidance strategy at the communication layer disclosed in this embodiment forms a perfect synergy and complement with the physical layer control in Embodiment 3. It demonstrates the multi-layered, adaptive resilience of the system in the face of complex park environments and control uncertainties. By using the diagnostic prediction results of the interpretable neural network XNN not only for root cause suppression at the physical layer but also for parameter feedforward adjustment at the communication layer, it ensures that the DC carrier communication link of the smart park maintains maximum possible availability and reliability under any operating condition, providing a solid communication foundation for realizing intelligent management and energy scheduling of the park.

[0250] In summary, Example 4 details how the system quickly identifies the failure through an intelligent judgment mechanism and immediately initiates a refined adaptive avoidance strategy at the communication layer when the physical layer power smoothing command fails to execute. This strategy proactively mitigates communication risks by locking onto the high-interference prediction state and actively adjusting key parameters such as modulation order, FEC coding redundancy, and communication bandwidth. The introduction of this mechanism makes the channel modeling and optimization system of this invention a complete solution with deep defense capabilities, greatly enhancing the system's practical value and robustness in real-world complex environments.

[0251] Example 5:

[0252] Corresponding to the above method embodiments, such as Figure 8 As shown, this invention also proposes a smart park multi-voltage level DC carrier adaptive channel modeling system, comprising:

[0253] The data acquisition module is used to collect DC carrier channel data within the smart park. The DC carrier channel data includes basic data for multiple voltage levels and distributed power source data. The multiple voltage levels include 10kV, 380V, and 220V voltage levels, and the distributed power sources include photovoltaic power sources or energy storage power sources.

[0254] The feature extraction module is used to extract four-path channel feature parameters based on the DC carrier channel data to form a feature set; wherein, the four paths include direct transmission path, load reflection path, environmental interference path, and distributed power source reflection path;

[0255] The channel modeling module is used to construct and apply an interpretable neural network XNN channel model based on the feature set through an edge-cloud collaborative approach. The XNN channel model has four sub-networks corresponding to the four paths respectively. Specifically, the channel modeling module is used to: train the XNN channel model on the cloud platform, compress the trained model and distribute it to the edge nodes, and the edge nodes load the compressed model to perform channel modeling inference.

[0256] An adaptive adjustment module is used to monitor the fluctuations of the distributed power source data. When the fluctuations meet preset adjustment conditions, the parameters of the XNN channel model are dynamically adjusted to achieve adaptive updates of the channel model.

[0257] This invention relates to a smart park multi-voltage level DC carrier adaptive channel modeling system. By constructing an innovative four-path channel model and combining it with an edge-cloud collaborative interpretable neural network (XNN) architecture, it achieves accurate, real-time modeling and adaptive tracking of complex DC carrier channels in smart parks. The core advantages of this solution lie in its interpretability and proactive closed-loop mechanism: First, the lightweight XNN engine deployed at the edge not only predicts channel conditions but also diagnoses the dominant interference paths leading to communication quality degradation in real time. Second, based on these interpretable diagnostic results, the system can proactively initiate preventative closed-loop optimization by sending precise control commands to the distributed power supply or load scheduling modules of the power physical layer, suppressing interference at its source. Furthermore, when physical layer control is obstructed due to priority conflicts, the communication layer adaptive avoidance strategy utilizes the XNN's prediction results to proactively adjust the communication transceiver parameters in a feedforward manner, ensuring the highest robustness and high availability of the communication link under complex power system operating conditions. This solves the problems of channel model mismatch, unclear interference sources, and the communication system's passive adaptation inherent in traditional solutions.

[0258] Example 6:

[0259] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0260] like Figure 9 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0261] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0262] Bus 102 may include a pathway for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0263] The memory 103 stores a computer program corresponding to a general page-turning data recursive query and processing method according to the above embodiments of the present invention. This computer program is controlled and executed by the processor 101. The processor 101 executes the computer program stored in the memory 103 to implement the content shown in the aforementioned method embodiments.

[0264] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 9 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0265] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for adaptive channel modeling of multi-voltage level DC carrier in smart parks, characterized in that, Includes the following steps: Collect DC carrier channel data within the smart park, including multi-voltage level basic data and distributed power source data; The four-path channel feature parameters of the DC carrier channel data are extracted to form a feature set; wherein, the four paths include the direct transmission path, the load reflection path, the environmental interference path, and the distributed power source reflection path; Based on the feature set, an interpretable neural network XNN channel model is constructed and applied through an edge-cloud collaborative approach. The XNN channel model has four sub-networks corresponding to the four paths respectively. The construction and application include: training the XNN channel model on the cloud platform, compressing the trained model and distributing it to the edge nodes, and the edge nodes loading the compressed model to perform channel modeling inference. The fluctuations in the distributed power source data are monitored, and when the fluctuations meet preset adjustment conditions, the parameters of the XNN channel model are dynamically adjusted to achieve adaptive updating of the channel model.

2. The method according to claim 1, characterized in that, The distributed power data includes power fluctuation parameters. and interface impedance parameters ; Among them, the The difference between the real-time grid-connected power and the rated power of the distributed power source; The This is the difference between the real-time impedance of the grid-connected interface of the distributed power source and the line reference impedance.

3. The method according to claim 2, characterized in that, The step of extracting the four-path channel feature parameters of the DC carrier channel data specifically includes extracting the feature parameters of the distributed power source reflection path, wherein the feature parameters include: Reflection power factor The calculation formula is as follows: ； in, Based on the fundamental reflection power factor; The rated power of the distributed power source; Impedance correction factor The calculation formula is as follows: ； in, The reference impedance of the line is denoted as .

4. The method according to claim 1, characterized in that, Training the XNN channel model on the cloud platform specifically includes: Define loss function The sum of the mean squared error (MSE) and the L1 regularization term, the The calculation formula is: ； in, The number of samples; These are the model's predicted values; These are measured values; These are the weights of the four sub-networks; The path correlation coefficient; and This is the regularization penalty coefficient.

5. The method according to claim 1, characterized in that, The edge nodes load the compressed model and perform channel modeling inference, specifically including: The compressed model loaded by the edge node is a lightweight inference engine that retains the core structure of the four sub-networks in the XNN channel model, which correspond to the direct transmission path, load reflection path, environmental interference path, and distributed power source reflection path, respectively. The edge nodes input the feature parameters collected in real time into the four sub-networks in parallel to obtain the channel influence components of the corresponding four paths. The four channel influence components are weighted and fused to obtain the final channel prediction result, thereby realizing interpretable real-time reasoning of the channel state at the edge.

6. The method according to claim 3, characterized in that, The preset adjustment condition is a composite statistical trigger condition, including at least one of the following: First preset time window Within, the power fluctuation parameters The moving average is greater than the first power average threshold; Second preset time window Within, the impedance correction factor The duration for which the impedance remains above the first impedance correction threshold is greater than a preset duration threshold. .

7. The method according to claim 6, characterized in that, The dynamic adjustment of the parameters of the XNN channel model specifically includes: According to the above The moving average or the Calculate a nonlinear adaptive adjustment factor based on the duration exceeding the threshold. ,in It is positively correlated with the moving average or the duration of exceeding the threshold; Apply the adaptive adjustment factor Amplify the output weights of the sub-networks corresponding to the distributed power source reflection paths in the XNN channel model; Meanwhile, the output weights of the other three sub-networks in the XNN channel model are normalized and reduced to maintain the stability of the model output.

8. The method according to claim 1, characterized in that, The method further includes performing preventative closed-loop channel optimization based on the interpretable output of the XNN channel model, specifically including: The edge node periodically performs channel modeling inference, and when the predicted channel communication quality parameters are about to deteriorate and reach the preset communication quality threshold, it triggers preventive optimization. The edge node parses the outputs of the four sub-networks of the XNN channel model and identifies the dominant path that is about to degrade the channel communication quality parameters; When the dominant path is the distributed power source reflection path, a power smoothing command is sent to the control module of the distributed power source to actively suppress power fluctuation parameters. fluctuation; When the dominant path is the load reflection path, a non-critical load time shift instruction is sent to the park load scheduling module to actively change the load reflection characteristics.

9. The method according to claim 8, characterized in that, The preventative closed-loop channel optimization also includes: After the edge node sends the power smoothing command to the control module of the distributed power source, a command execution confirmation timer is started. If, before the timer expires, the control module returns a command to refuse the response, or if no execution confirmation response is received and the control module detects that the timer has expired, then the control module has refused the response. If the fluctuation does not converge as expected, the power smoothing instruction is deemed to have failed, and the communication layer adaptive avoidance strategy is immediately activated. The adaptive avoidance strategy at the communication layer specifically includes: The output of the sub-network corresponding to the distributed power source reflection path in the XNN channel model is locked as a high-interference prediction state; Based on the high interference prediction state, the edge node actively sends a communication parameter pre-adjustment command to the transceiver carrying the DC carrier communication. The communication parameter pre-adjustment command includes at least one of the following: Reduce the modulation order of modulation and demodulation; Increase the coding redundancy level of Forward Error Correction (FEC); Switch the communication subcarrier to the spare frequency band that has the least impact from the high interference prediction state.

10. A smart park multi-voltage level DC carrier adaptive channel modeling system, characterized in that, include: The data acquisition module is used to collect DC carrier channel data within the smart park. The DC carrier channel data includes basic data for multiple voltage levels and distributed power source data. The multiple voltage levels include 10kV, 380V, and 220V voltage levels, and the distributed power sources include photovoltaic power sources or energy storage power sources. The feature extraction module is used to extract four-path channel feature parameters based on the DC carrier channel data to form a feature set; wherein, the four paths include direct transmission path, load reflection path, environmental interference path, and distributed power source reflection path; The channel modeling module is used to construct and apply an interpretable neural network XNN channel model based on the feature set through an edge-cloud collaborative approach. The XNN channel model has four sub-networks corresponding to the four paths respectively. Specifically, the channel modeling module is used to: train the XNN channel model on the cloud platform, compress the trained model and distribute it to the edge nodes, and the edge nodes load the compressed model to perform channel modeling inference. An adaptive adjustment module is used to monitor the fluctuations of the distributed power source data. When the fluctuations meet preset adjustment conditions, the parameters of the XNN channel model are dynamically adjusted to achieve adaptive updates of the channel model.

Images