A system and method for predicting optical power of distribution network optical cables

Abstract

This invention discloses a system and method for predicting optical power in distribution network optical cables, relating to the field of power optical communication technology. It addresses problems such as high misjudgment rate, prediction lag, and the tendency of traditional optimization algorithms to get trapped in local optima in optical power prediction for distribution network optical cables. The invention proposes an IALA-TCN-BiLSTM fusion prediction model, which enhances population diversity through an averaging search mechanism and accelerates convergence by combining a Lévy flight strategy with dynamic threshold truncation. Multi-scale spatiotemporal features of optical power extracted from the TCN network are input into the BiLSTM network to model long-period attenuation patterns, and optical switch polling technology is integrated to achieve multi-channel optical cable data acquisition. The prediction model constructed by this invention can optimize the operating status of optical cables, enabling advanced early warning of optical power degradation trends in distribution network topologies. When applied to power communication optical cable monitoring systems, it can significantly improve fault prediction accuracy and system stability, providing technical support for intelligent operation and maintenance of distribution networks.

Description

Technical Field

[0001] This invention relates to the field of power optical communication technology, specifically to a system and method for predicting the optical power of distribution network optical cables. Background Technology

[0002] With the continuous advancement of intelligent transformation of power distribution networks, the scale of all-dielectric self-supporting optical cables (ADSS) laid along power lines is rapidly increasing, forming a backbone architecture for power distribution communication covering urban and rural areas. However, optical cables have long faced severe challenges in complex line corridor environments, and the existing monitoring system suffers from three major defects: First, it relies on a passive response mechanism of communication network management status polling, which can only passively detect anomalies through offline equipment alarms and cannot proactively predict degradation trends; second, it lacks fault identification capabilities, making it difficult to distinguish between different fault types such as fiber core breakage, water ingress into junction boxes, optical module failure, or terminal power failure; and third, it has low handling efficiency, requiring 2-3 teams of personnel to conduct OTDR testing and on-site investigation after a fault occurs, with an average investigation time exceeding 4 hours, and the preset threshold alarm method cannot adapt to dynamic changes in optical power, leading to a significant increase in maintenance costs. The core problem lies in the current system's lack of real-time modeling and prediction capabilities for optical cable attenuation trends. Traditional time-series optimization algorithms are prone to getting trapped in local optima and are unable to capture the multi-scale characteristics and long-cycle attenuation patterns of optical power data, urgently requiring innovative prediction methods to overcome technical bottlenecks.

[0003] Therefore, how to improve the ability to predict potential hazards in advance to identify and eliminate potential fault hazards and ensure the continuous and reliable operation of communication links is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention discloses a power prediction system and method for distribution network optical cables. By establishing a power prediction model, the system dynamically senses and predicts the operating status of multiple optical cables. Based on time-series data of independent optical cables collected through optical switch polling, and combining an improved Artificial Lemming Algorithm (IALA) and a TCN-BiLSTM fusion prediction model, it extracts local power fluctuation characteristics and long-period attenuation patterns, achieving early warning of optical cable degradation trends. This enables intelligent management and control of the entire distribution network optical cable chain, from data acquisition and degradation prediction to adaptive regulation, improving the fault prediction capability and operational reliability of power optical communication systems.

[0005] A power prediction system for distribution network optical cables includes an OTDR module, a central processing unit module, an optical switch module, a photoelectric conversion module, and a communication module; the central processing unit module is electrically connected to the OTDR module, the optical switch module, the photoelectric conversion module, and the communication module.

[0006] The OTDR module is used to transmit probe light pulses to the optical cable line and receive the echo light signals returned from the optical cable to obtain the loss characteristics along the optical cable path, which is used to determine the location and type of the fault point and the optical power distribution of the optical cable under test.

[0007] The optical switch module is used to automatically switch between multiple optical cables, enabling a single OTDR module to poll and detect multiple optical cable lines, thereby achieving online monitoring of multiple channels;

[0008] The photoelectric conversion module is used to convert the echo optical signal received by the OTDR module into an electrical signal and transmit it to the central processing unit module for subsequent analysis and processing.

[0009] The communication module is used to transmit the processing results and prediction data to a remote server, monitoring platform or host computer system via wired or wireless means, so as to realize remote data uploading, instruction receiving and remote control functions;

[0010] The central processing unit module is used for high-speed real-time data processing, feature extraction and fault identification of the electrical signals converted by the photoelectric conversion module. At the same time, it integrates the IALA-TCN-BiLSTM prediction model to realize intelligent prediction of the trend of optical power change.

[0011] This invention also provides a method for predicting the optical power of distribution network optical cables, which is implemented through the aforementioned prediction system; the implementation steps of the prediction method are as follows:

[0012] Step S1: Use the OTDR module and optical switch module to switch the optical path to collect data on the optical power of each optical cable in the distribution network;

[0013] Step S2: Preprocess the collected optical power data of each optical cable in the distribution network to obtain preprocessed optical power time series data;

[0014] Step S3: Use TCN network to extract features from optical power time series data, obtain multi-scale time-dependent features by stacking causal dilated convolution, introduce multi-layer residual block structure, and enhance model stability by residual skip connection and Dropout regularization after the output of each dilated convolution layer.

[0015] Step S4: Construct an IALA-TCN-BiLSTM fusion prediction model. The optical power temporal features extracted by the TCN network are used as input, and the long-short-term dependencies of the temporal features are modeled by BiLSTM. The output of the fusion prediction model is iteratively optimized by the error between the loss function and the true value, and the hyperparameters of the fusion prediction model are optimized by the IALA algorithm to achieve training of the fusion prediction model.

[0016] Step S5: The trained IALA-TCN-BiLSTM fusion prediction model predicts the real-time acquired optical power data. The fusion prediction model sequentially completes feature extraction, sequence modeling, and prediction output to obtain the optical power trend results within the target time window.

[0017] The beneficial effects of this invention are as follows: This invention discloses a method for predicting optical power in optical cables by deeply integrating adaptive learning rate optimization and spatiotemporal features. It uses the improved Artificial Lemming Algorithm (IALA) to perform global hyperparameter optimization on the TCN-BiLSTM hybrid model. By deeply integrating the efficient local feature extraction capability of the Temporal Convolutional Network (TCN) with the temporal dependency modeling advantage of the Bidirectional Long Short-Term Memory Network (BiLSTM), a dynamic prediction model for optical power is constructed.

[0018] This method offers an innovative solution to address the challenges of traditional time-series models (such as single LSTM) failing to capture the multi-scale local fluctuations in optical power data, traditional optimization algorithms (such as Adam) easily getting trapped in local optima leading to slow model convergence, and the insufficient interpretability of purely data-driven models. Its innovation is primarily reflected in the model structure: the dilated causal convolutional layer of TCN complements the bidirectional time-series modeling of BiLSTM, balancing the capture of local abrupt changes with the fitting of long-term trends.

[0019] Secondly, at the algorithm optimization level: the improved IALA algorithm dynamically adjusts the learning rate by maintaining gradient direction consistency, and combined with the Halton sequence initialization strategy, significantly improving the efficiency of hyperparameter search and the model generalization ability;

[0020] At the engineering application level: the preprocessing module integrating wavelet denoising and moving average effectively suppresses impulse noise and random interference in optical power data, enhancing the robustness of the model in complex real-world environments. By deploying this method in a power communication optical cable monitoring system, core functions such as early warning of abnormal optical power fluctuations and location of fiber optic link attenuation faults can be achieved, assisting maintenance personnel in proactively intervening in potential faults and reducing the risk of communication interruptions caused by optical cable performance degradation. This provides reliable technical support for the intelligent operation and maintenance and condition assessment of power communication optical cables, demonstrating significant economic benefits and practical engineering value. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1This is a block diagram of a power prediction system for distribution network optical cables according to the present invention.

[0023] Figure 2 This is a structural diagram of the hardware device in the optical power prediction system for distribution network optical cables according to the present invention.

[0024] Figure 3 This is a flowchart of a method for predicting optical power in a distribution network optical cable according to the present invention;

[0025] Figure 4 This is a flowchart of the hyperparameter optimization of the IALA-TCN-BiLSTM fusion prediction model using the IALA algorithm. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Specific Implementation Method 1: Combination Figure 1 and Figure 2 This embodiment describes a power prediction system for distribution network optical cables. The system includes a power supply module, an OTDR module, a central processing unit module, an optical switch module, a photoelectric conversion module, and a communication module. The central processing unit module is electrically connected to the OTDR module, the power supply module, the optical switch module, the photoelectric conversion module, and the communication module.

[0028] The power supply module provides a stable power supply to the OTDR module, central processing unit module, optical switch module, photoelectric conversion module, and communication module in the prediction system. The power supply module includes a voltage regulator circuit and an overcurrent protection circuit to ensure the reliability and safety of the system's long-term continuous operation. Figure 2 As shown, in this embodiment, the power supply module uses an LM317 three-terminal regulator to form a voltage regulation circuit. The input voltage is 24VDC, and the output voltage is adjusted to 5V DC and 3.3V DC through voltage divider resistors, which power the central processing unit module and the OTDR module respectively. The overcurrent protection circuit consists of a self-resetting fuse and a TVS diode connected in parallel. When the current exceeds the threshold, the fuse is triggered to open the circuit, and the transient surge voltage is absorbed by the TVS diode.

[0029] The OTDR module is used to transmit probe light pulses to the optical cable line and receive backscattered or reflected signals returned from the optical cable, thereby obtaining the loss characteristics along the optical cable path, which is used to determine the location and type of the fault point and the optical power distribution of the optical cable under test.

[0030] The optical switch module is used to automatically switch between multiple optical cables, enabling a single OTDR module to poll and detect multiple optical cable lines, realize online monitoring of multiple channels, and greatly improve the applicability and monitoring efficiency of the system.

[0031] The photoelectric conversion module is used to convert the echo optical signal received by the OTDR module into an electrical signal and transmit it to the central processing unit module for subsequent analysis and processing. This module has high sensitivity and low noise characteristics to ensure detection accuracy.

[0032] The communication module is used to transmit processing results and prediction data to a remote server, monitoring platform, or host computer system via wired or wireless means, realizing remote data uploading, command reception, and remote control functions. The communication module supports multiple communication protocols to adapt to different application environments.

[0033] The central processing unit module is used for high-speed real-time data processing, feature extraction, and fault identification of the echo signals acquired by the OTDR module. It also integrates the IALA-TCN-BiLSTM fusion prediction model to achieve intelligent prediction of optical power variation trends. The specific process is as follows:

[0034] Collect raw data on the optical power of each optical cable in the power distribution network;

[0035] The collected raw data is preprocessed to obtain the preprocessed optical power data sequence of each optical cable, and the dataset is then divided.

[0036] TCN network is used to extract the temporal features of optical power, and ReLU laser function and Dropout regularization are combined to enhance nonlinear representation and generalization ability.

[0037] An IALA-TCN-BiLSTM fusion prediction model was constructed, and the IALA algorithm was introduced to optimize the model's learning rate and number of neurons, thereby improving prediction accuracy and generalization ability.

[0038] The preprocessed optical power time series is input into the trained IALA-TCN-BiLSTM fusion prediction model, which outputs the optical power trend within the target time period.

[0039] In this embodiment, the method also includes real-time acquisition of optical power data of each optical cable in the distribution network. The optical power data of each optical cable is collected in real time by controlling the optical switch. During the acquisition process, the central processing unit module switches the optical switch channel sequentially according to the set polling cycle, so that the OTDR module can accurately obtain the current optical power information of each optical cable by channel, ensuring the integrity and real-time performance of data acquisition, and providing reliable input data support for subsequent trend prediction.

[0040] In this embodiment, the OTDR module includes a 1550nm laser diode and an InGaAs photodiode. The laser pulse is directionally emitted to the optical cable under test via a circulator, and the backscattered signal is transmitted to the photoelectric conversion module through the circulator port. The module incorporates an XC7Z020 FPGA to control the pulse emission timing and has a pre-set Rayleigh scattering and Fresnel reflection signal feature library to achieve real-time analysis of the optical cable loss curve.

[0041] In this embodiment, the central processing unit (CPU) module is built on an ARM Cortex-A72 processor. It preprocesses the OTDR echo signal using a wavelet denoising algorithm, extracts the temporal features of optical power using a TCN convolutional network, and captures the long-short-term dependencies in the signal using its dilated causal convolutional layers. Subsequently, the temporal features are input into the ALA-TCN-BiLSTM prediction model. The model parameters are trained using historical optical power data, and the learning rate is optimized using the Adam optimizer to achieve intelligent prediction of optical power change trends. The trained model is then saved to the CPU module.

[0042] In this embodiment, the optical switch module includes a 1×8 MEMS optical switch array, and polling commands are sent by the central processing unit module through an RS485 interface. Each channel optical cable interface uses an FC / APC connector and is configured with an optical power threshold detection circuit to complete the function of polling for optical cable degradation in the distribution network.

[0043] In this embodiment, the photoelectric conversion module uses an APD diode as the light receiving device. Its output current is converted into a voltage signal by a transimpedance amplifier and then quantized by a 24-bit ADC. Noise suppression is achieved through a two-stage RC low-pass filter to ensure a dynamic range >60dB. The converted digital signal is transmitted to the central processing unit module via an SPI interface.

[0044] In this embodiment, the communication module integrates an Ethernet interface, a Type-C interface, and an RS485 bus interface; the communication module integrates a 4G LTE module and dual-band Wi-Fi, and also provides an RJ45 gigabit Ethernet interface to support adaptive access in multiple network environments; the data transmission protocol uses the MQTT protocol to push real-time alarm information, uses the FTP protocol to complete periodic full data uploads, and receives remote control commands based on the Modbus TCP protocol.

[0045] Specific Implementation Method Two: Combination Figure 3 and 4This embodiment describes a prediction method for a distribution network optical cable optical power prediction system as described in Specific Embodiment 1. To improve the intelligence level of this invention in optical cable fault early warning and trend analysis, the central processing unit module preferably integrates an IALA-TCN-BiLSTM fusion prediction model for high-precision modeling and trend prediction of the collected historical optical power time-series data. The specific process is as follows:

[0046] Step S1: Data collection of optical power of each optical cable in the power distribution network is achieved by switching the optical path through the OTDR module and optical switch;

[0047] Step S2: Preprocess the collected optical power data of each optical cable in the distribution network to obtain preprocessed time series data of each optical cable. Divide the data into training, validation, and test sets according to time proportions for model training, parameter tuning, and performance evaluation; the specific process is as follows:

[0048] In this embodiment, the data preprocessing methods mainly include noise reduction, smoothing, and data normalization.

[0049] The specific steps of the noise reduction process include: dynamically identifying the starting boundary of the optical fiber end reflection noise based on the sliding window variance algorithm, and truncating the tail data sequence whose variance exceeds a preset threshold after the window, so as to eliminate invalid fluctuation interference caused by optical fiber endface echo in the optical power signal; in this embodiment, the accurate determination of the effective data endpoint is achieved through adaptive noise boundary localization technology, as shown in the following formula:

[0050]

[0051] Where, x i is the optical power value at the i-th distance point; w is the width of the sliding window; is the window mean; S is the total number of sampling points for the optical power sequence of the tested optical cable. The truncation point is selected as follows: when... Cut off at time, among which Let Variance be the optical power variance of the q-th sliding window; α represents the variance of the historical maximum optical power; α is the noise filtering threshold coefficient, α∈(0,1];

[0052] The data smoothing process specifically includes employing the Savitzky-Golay time-domain filtering algorithm, based on the principle of polynomial least squares fitting. This algorithm dynamically models the local features of the optical power signal within a sliding window and performs low-distortion noise suppression, as shown in the following formula:

[0053] A polynomial is used to fit the local window data:

[0054]

[0055] Among them, y i x is the smoothed value of the i-th distance point in the output sequence. i+j Given the input sequence of (i+j) original sampled values, ν = (w-1) / 2 is the half-window width, and c j The convolution coefficients are determined by the polynomial order d and the sliding window width w.

[0056] The normalization process refers to independently normalizing all historical data for each optical cable, using the following formula:

[0057]

[0058] Where x is the original optical power sample value, x norm The standardized optical power value, μ valid σ is the mean of the valid data segment before truncation. valid The standard deviation of the effective data segment is used, and the preprocessed data is divided into training set, validation set and test set according to the time ratio for model training, parameter tuning and performance evaluation.

[0059] Step S3: Feature extraction is performed on the preprocessed optical cable time series data in step S2. A temporal convolutional neural network (TCN) is used to obtain multi-scale time-dependent features by stacking causal dilated convolutions. A residual structure is introduced to avoid gradient vanishing. The ReLU activation function and Dropout regularization are combined to improve the nonlinear expressive power and generalization performance of the model.

[0060] In this embodiment, time-series features of optical power in the distribution network are extracted using a temporal convolutional neural network (TCN) for multi-scale feature extraction. The specific implementation process is as follows: The normalized time-series data of each optical cable is truncated by a fixed time window length k to generate multiple training samples. Each sample consists of data from k consecutive time window points. The input data construction formula is as follows:

[0061]

[0062] Where M = N - k + 1 is the number of samples, and N is the normalized data length. Zero padding is performed inside the convolutional layer to ensure temporal causality.

[0063] To capture the dependencies across different time spans in optical power time-series data, a causal dilated convolution is constructed to expand the receptive field through hierarchical dilation coefficients. A four-layer TCN residual block is built, with each layer extracting features across different time spans through dilated convolution. The convolution operation formula for the l-th layer is as follows:

[0064]

[0065] in, This indicates that the l-th convolutional kernel applies to historical time-series data X. (t) Multi-scale feature extraction results; expansion coefficient d l =2 l-1 As the number of layers increases exponentially, the convolutional kernel moves at a rate of d on the time axis. l Historical data is sampled at intervals. A causal convolutional kernel with a width of K=5 ensures that it depends only on the current and past inputs, while the number of output channels increases exponentially with each layer, controlled by weights. To achieve feature dimension expansion, where China F l-1 F represents the number of feature dimensions in the (l-1)th layer. l The number of feature dimensions in layer l enables the lower layers to capture both reflective and non-reflective events, forming a multi-scale feature fusion capability. l This is a bias term.

[0066] After each layer of dilated convolution output, residual skip connections and Dropout regularization are used to enhance model stability. The specific implementation steps are as follows:

[0067] The residual connection structure uses the output of each layer of dilated convolution to... With the original input X (t) The addition of Dropout regularization, while preserving the original temporal information, suppresses overfitting and alleviates the gradient vanishing problem. Identity mapping ensures the stability of deep network training, Dropout randomly masks 20% of neurons to enhance generalization ability, and temporal alignment constraints guarantee consistent feature propagation. The specific formulas are as follows:

[0068]

[0069] The Dropout operation uses δ l Randomly masking 20% ​​of neuron outputs suppresses overfitting of the model to the training data; directly passing the original input to the output ensures stable gradient propagation and alleviates the degradation problem of deep networks.

[0070] After completing the multi-level residual convolution feature extraction, global average pooling is used to compress the temporal dimension, generating a low-dimensional, high-information-density feature vector. The specific implementation steps are as follows:

[0071] The temporal feature tensor output by the fourth-layer TCN is processed by global average pooling. Compression is performed along the time dimension to generate a 256-dimensional feature vector. The calculation formula is:

[0072]

[0073] Among them, F outThis represents the output feature vector after global average pooling, generated by averaging the residual output of the fourth-layer TCN along the time dimension. It corresponds to the number of channels in the last layer of the TCN network, encoding key features of reflection events such as local abrupt changes in optical power signals and non-reflection events such as fiber splicing, providing high-information-density input for subsequent BiLSTM prediction modules. This represents summing the residual outputs of all time steps within the time window and dividing by k to achieve global average pooling. `g` is the time window index, and `k` represents the time window length. A sliding window is used to truncate the input sequence, ensuring the model perceives the complete optical power fluctuation cycle. The residual connection design improves training convergence speed while ensuring temporal causality. The calculation process is as follows:

[0074]

[0075] In the formula, W4 is the convolution kernel weight matrix, b4 is the bias vector, and X... (t-8g) X represents the sequence of input signals within a specific time window. (t-k+g) This indicates a residual connection, which allows the original input to be directly superimposed on the output, preserving the short-term fluctuation characteristics of optical power.

[0076] Step S4: Construct an IALA-TCN-BiLSTM fusion prediction model. The temporal features of optical power extracted by the TCN network are used as input, and the long-short-term dependencies are modeled using BiLSTM. The model output is iteratively optimized using the error between the loss function and the true value to improve prediction accuracy. Furthermore, the IALA algorithm is introduced to adjust key hyperparameters of the model, mainly including the learning rate and the number of BiLSTM neurons, thereby improving model performance and generalization ability.

[0077] In this embodiment, during the training of the fusion prediction model, the IALA algorithm specifically introduces a Logistic chaotic mapping to initialize the population, uses the weighted average absolute percentage error (WMAPE) as the fitness value, and globally optimizes the key-value vector dimension of the dilation factor of the temporal convolutional network, the number of residual blocks, and the inter-layer dropout rate of the bidirectional long short-term memory network in the IALA-TCN-BiLSTM model. For example... Figure 4 As shown:

[0078] S41. The TCN network is used to extract multi-scale local features through causal dilated convolution, BiLSTM is used to model long temporal dependencies, and residual connections are used to prevent gradient vanishing, forming a spatiotemporal feature fusion network architecture.

[0079] Step S42: Initialize the weights of the convolutional kernels in the TCN network using the He normal distribution, the bias of the BiLSTM input / forget gate, and orthogonally initialize the parameters of the recurrent layers to ensure the stability of feature propagation and the controllability of gradients; the initialization formula is:

[0080]

[0081] in, This represents the number of input channels in the l-th convolutional layer. The formula for the first TCN convolutional kernel is: d embed For the embedded dimension.

[0082] The BiLSTM parameter initialization formula, which employs orthogonal initialization to enhance temporal feature propagation, is as follows:

[0083] W lstm =QΛQ T Q T Q = I (10)

[0084] Where Q is an orthogonal matrix, whose distance-preserving and gradient-stability properties enhance the temporal feature transfer capability of BiLSTM and improve model performance. I is the identity matrix, and Λ is a diagonal matrix with elements following a uniform distribution. The input gate and forget gate biases are initialized to a vector of all 1s: b f ,b i =1. The bias term initialization implements the initialization strategy for bias vectors in the network layers. Specifically, the biases of fully connected layers are zero-initialized: b fc =0, the convolutional layer bias is initialized with a constant, and the output variance of each layer is ensured to be consistent through layer-by-layer scaling. The specific formula is:

[0085] Step S43: Because the traditional artificial lemming algorithm relies too heavily on the unidirectional guidance of the optimal individual, the population rapidly clusters in local optima. The linear difference pattern causes the search directions to converge, deteriorating the population diversity index and restricting the global optimization capability. Therefore, the improved IALA algorithm is adopted to initialize the population through Logistic chaotic mapping, integrate the averaging search strategy and truncated Levy flight, and dynamically optimize the expansion coefficient of TCN and the hyperparameters of BiLSTM.

[0086] In this embodiment, the specific steps of the hyperparameter optimization process of the IALA algorithm for the IALA-TCN-BiLSTM fusion prediction model are as follows:

[0087] Step S431: Introduce an averaging search mechanism to address the problems of insufficient population diversity and singular local search direction caused by over-reliance on the unidirectional guidance of the optimal individual. The formula is as follows:

[0088]

[0089] in, Z represents the average position in the γ-th dimension, generated by fusing the position values ​​of the current individual ξ and the random non-optimal individual r in that dimension.ξ,γ Z represents the position value of the current individual ξ in the γ-th dimension, indicating the state of the individual in the current dimension during the optimization process. r,γ This represents the position value of a randomly selected non-optimal individual r in the γ-th dimension. The variance of the average position reflects the stability of the dimensional values ​​after fusion. This search mechanism breaks the original algorithm's unidirectional guidance mode that relies solely on the most abundant individual, preventing the population from rapidly clustering in locally optimal regions and significantly enhancing population diversity.

[0090] Let the set of hyperparameters be:

[0091] θ=(η,B,H,δ l (13)

[0092] Where η is the learning rate, B is the batch size, defines the number of samples for a single training iteration, H is the number of hidden layer nodes, and δ l This is the Dropout rate.

[0093] Set the optimization objective function, and define the fitness function as the negative RMSE of the validation set, as shown in the following formula:

[0094]

[0095] Define the search space: The optimization problem is π = 4-dimensional, corresponding to 4 hyperparameters, and the search range for each dimension γ is [LB]. γ UB γ ], among which, LB γ UB represents the lower bound of the γ-th dimension. γ It represents the upper bound of the γ-th dimension.

[0096] Step S432: Randomly generate the initial population: The population consists of N individuals, each individual's position... Randomly initialize in d-dimensional space; the formula is:

[0097] z ξ,γ =LB γ +rand·(UB γ -LB γ ),ξ=1,2,...,N; γ=1,2,...,π (15)

[0098] Among them, z ξ,γ Let ξ represent the position of the ξ-th individual in the γ-th dimension, and rand be a random number uniformly distributed in [0,1]. The positions of all individuals form an N×d matrix. The formula is as follows:

[0099]

[0100] Step S433: Iteratively update IALA. At the beginning of each iteration, calculate the energy factor E(t) to control the balance between exploration and development, as shown in the following formula:

[0101]

[0102] Among them, T max is a constant, representing the maximum number of iterations, and rand is a uniformly distributed random number.

[0103] Step S434: In the exploration phase, when E(t) > 1, i.e. long-distance migration behavior, the global diversity is enhanced by combining the averaging search mechanism, i.e., by enhancing global diversity through the following formulas (18) and (19); the specific formulas are expressed as follows:

[0104]

[0105] in, Let ξ be the position vector of the ξ-th lemming individual at iteration step t. The position is the historical best position found in the population up to generation t, and F is the direction control factor. For Brownian motion random vectors, Let L be the mixed weight vector, and L be the adaptive step size factor. The average position of the γth dimension

[0106] When the energy factor E(t) ≤ 1, it is considered a hole-digging behavior. We introduce truncated Lévy flight to improve local search efficiency. The formula for truncated Lévy flight is as follows:

[0107]

[0108] Where θ is the scale parameter, controlling the shape of the Levy step size distribution, Levy(θ) is the original Levy distribution, and G is the dynamic cutoff coefficient. The cutoff boundary is: G = G0·(1-t / T) max G0 is the initial threshold, and sign(·) is the sign function, which means that the original step direction is retained after truncation to avoid invalid oscillations.

[0109] In the IALA algorithm, this threshold is related to the number of iterations, ensuring the algorithm maintains strong exploratory capabilities. In the improved Artificial Lemming Algorithm (IALA), the individual's position update uses a truncated Lévy step size; the improved formula is as follows:

[0110]

[0111] in, This represents the historical best position in the t-th iteration. This represents the historical best position in the (t+1)th iteration. Let be the position vector of individual i at the t-th iteration. G is a dynamic truncation coefficient that gradually decreases with the number of iterations, limiting the maximum value of the Lévy step size. When the generated Lévy step size exceeds the threshold of G, the actual step size is limited to G to avoid invalid large-scale jumps. In the above formula, the truncation of the Lévy distribution is achieved in the following way:

[0112] In the dynamic truncation threshold setting, G serves as the dynamic threshold, displaying the maximum value of the Lévy step size. In the early stages, a larger G allows for a larger Lévy step size, enhancing global exploration. Later, G approaches 0, limiting the step size to small perturbations and emphasizing local exploration.

[0113] Step S435: Perform population update, i.e., when the individual parameters exceed the preset search range [LB] after the update. γ UB γ When an out-of-bounds individual is found, a reflection correction strategy is used to pull it back into the feasible region, avoiding the generation of invalid solutions. The reflection correction formula for out-of-bounds individuals is as follows:

[0114]

[0115] Among them, z ξ,γ Let LB represent the position value of the ξ-th lemming individual in the γ-dimensional parameter space. γ UB is the lower bound of the γ-th dimension parameter. γ This is the upper bound of the γ-th dimension parameter.

[0116] If the value still exceeds the bounds after reflection, it is forcibly assigned the boundary value, where clip(·) is the boundary truncation function, as shown in the following formula:

[0117] z ξ,γ =clip(z) ξ,γ ,LB γ UB γ ) (twenty two)

[0118] In this implementation, to avoid losing outstanding individuals during iteration, an elite retention mechanism is adopted, specifically global optimal retention and local elite injection:

[0119] Global optimal solution retention, always retaining the historical optimal solution. It is prohibited from participating in mutation operations, among which The formula for the new optimal individual generated in the current iteration is as follows:

[0120]

[0121] in, This represents the historical best position in the t-th iteration. This represents the historical best position in the (t+1)th iteration. For the newly generated candidate solutions, f(·) is the fitness function.

[0122] Local elite injection: Every 5 iterations, an elite individual is randomly selected from the historical population to replace the individual with the worst fitness in the current population, preventing population degradation. The formula is as follows:

[0123]

[0124] Where Population represents the current population set. This is the individual with the worst fitness in the current population. For historical elite individuals.

[0125] Step S436: Save the current optimal solution This solution represents the optimal combination of parameters for the output.

[0126] Step S44: During training, the weighted average absolute percentage error (WMAPE) is used as the fitness function. The Adam optimizer updates parameters through backpropagation, the early stopping mechanism monitors and validates the loss, and IALA dynamically adjusts the learning rate and dropout rate to improve convergence efficiency, thus obtaining the trained IALA-TCN-BiLSTM fusion prediction model. The specific implementation process is as follows:

[0127] The training set is input into the IALA-TCN-BiLSTM fusion prediction model, and the IALA algorithm is used to determine the optimal parameter combination of the IALA-TCN-BiLSTM model. The specific implementation steps are as follows:

[0128] The IALA algorithm is used to globally optimize four key hyperparameters of the IALA-TCN-BiLSTM model, establishing the following four-dimensional search space:

[0129]

[0130] Where η is the learning rate, B is the batch size (i.e., the number of samples in a single training iteration), H is the number of hidden layer nodes, and δ l This is the Dropout rate.

[0131] The optimal hyperparameters are determined using the IALA algorithm, after N iterations T max After optimization, the set θ with the highest fitness is selected. * , that is, θ * =(η * B * H * ,δ l * ) as the optimal parameter combination, where η * B * H * δ lThese correspond to η, B, H, and δ, respectively. l * The optimal hyperparameters; during the model training phase, the hyperparameters (η) optimized based on IALA. * B * H * ,δ l * Constructing a hybrid architecture: The TCN module adopts a hierarchical expansion factor d l =2 l-1 Causal convolution, with the number of residual blocks in N res =[H *

[64] Dynamically adjust and configure three sets of residual convolutional layers with gradually increasing channel numbers to achieve multi-scale feature extraction; the BiLSTM module sets the bidirectional hidden layer dimension H. ＊ Interlayer Dropout rate δ l * To prevent overfitting, the Adam optimizer is used for parameter updates. Temporal data is loaded in batches B. During forward propagation, local features are first extracted using a TCN, and then temporal dependencies are modeled using a BiLSTM. A delta parameter is applied to the fully connected layer. l * The predicted value is output after regularization. The training process uses MSE as the loss function, and backpropagation is performed according to η. ＊ The update step size is adjusted, and early stopping is triggered when the rate of change of the validation loss is less than 1 for five consecutive epochs, saving the optimal parameters and completing training. This process effectively controls model complexity while ensuring feature extraction capabilities through dynamic parameter configuration and regularization strategies.

[0132] Step S45: Training complete. Save the parameter combinations such as the number of TCN layers, the number of hidden nodes in BiLSM, and the Dropout rate after IALA optimization, solidify the IALA-TCN-BiLSTM fusion prediction model, and deploy it to the central processing unit module.

[0133] Step S5: The trained IALA-TCN-BiLSTM fusion prediction model predicts optical power data in real time. The fusion prediction model sequentially completes feature extraction, sequence modeling, and prediction output to obtain the optical power trend results within the target time window; the specific implementation process is as follows:

[0134] The preprocessed optical power time series validation set data is input into the trained IALA-TCN-BiLSTM fusion prediction model to output the optical power prediction trend for the future target time.

[0135] The preprocessed optical power time series of each root is input into the optimized IALA-TCN-BiLSTM fusion prediction model to perform multi-step prediction and generate the final prediction sequence. The specific steps are as follows:

[0136] Historical optical power data is extracted into column vector form according to time windows and used as the input sequence X. t-k:t The construction formula is as follows:

[0137]

[0138] Where k is the length of the input time window, t is the current time point corresponding to the feature of the acquired optical power data, tk is the starting time point of the window, tk:t is the time window range, and X t X is the optical power value at the end of the window. t-k This represents the optical power value at the starting point of the window. The input sequence is fed into the IALA-optimized TCN-BiLSTM model, which outputs the predicted future optical power value, as shown in the following formula:

[0139]

[0140] Where t+1 is the prediction start time, τ is the prediction step size, t+τ is the prediction end time, and t+1:t+τ is the prediction time window range. θ * =(η * B * H * ,δ l * () represents the optimal parameter set optimized by the IALA algorithm, and the final output sequence of the model is The final prediction results for each optical cable are obtained.

[0141] Based on the prediction results, the corresponding evaluation indicators are calculated, including mean absolute error (MAE), mean squared absolute error (RMAE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 The corresponding formulas are as follows:

[0142]

[0143] Among them, y i Let i be the actual value of the i-th sample. Let be the predicted value for the i-th sample. is the average of all samples, and n is the total number of samples.

[0144] Determine if the evaluation metrics meet the training threshold; if so, complete the training of TCN-BiLSTM; otherwise, retrain. The smaller the RMSE and MAE values, the better the model's predictive performance; R... 2 ∈[0,1], R 2 The closer a value is to 1, the better the fit, and the closer the predicted value is to the actual observed value.

[0145] The IALA-TCN-BiLSTM model constructed in this invention enables real-time degradation prediction of optical cables in distribution networks and accurately identifies abnormal conditions. The model integrates TCN multi-scale feature extraction with BiLSTM time-series modeling, optimizes hyperparameters through the IALA algorithm, and improves reliability by combining sliding window variance and filtering denoising. Based on optical switch polling, it provides early warning of degradation trends. This enables intelligent management and control of power optical communication systems, significantly improving stability and self-healing capabilities, and supporting the intelligent upgrading of distribution networks.

[0146] The various implementation examples described in this invention are analyzed layer by layer through a progressive architecture. The design essence of each example focuses on highlighting its unique innovative technical elements, while common implementation elements can be obtained through cross-referencing the preceding clauses. For implementation examples involving equipment components, based on the mapping relationship between their design logic and the previous process-type examples, this invention will focus on a general description of the system architecture. Detailed operational content can be found by referring to the technical analysis of the corresponding process-type examples, thereby achieving an organic coordination between the systematic nature of the technical explanation and the efficiency of the writing.

[0147] The detailed explanation of the disclosed technical solutions aims to enable those skilled in the art to accurately implement or apply this innovative achievement. Adjustments and improvements to these technical solutions are technically feasible for those with professional backgrounds. The core technical principles described in this invention can be extended to other technical forms for concretization, provided that the essence of the innovation and application boundaries are not deviated from. Therefore, the scope of protection of this invention should not be limited to the embodiments shown in the document, but should cover all applicable fields expanded based on the innovative principles and technical features, the boundaries of which are jointly defined by the technical essence and innovative value set forth in the claims.

Claims

1. A method for predicting optical power in distribution network optical cables, characterized by: This method is implemented through a prediction system; the system includes an OTDR module, a central processing unit module, an optical switch module, a photoelectric conversion module, and a communication module; the central processing unit module is electrically connected to the optical switch module, the photoelectric conversion module, and the communication module. The OTDR module is used to transmit probe light pulses to the optical cable line and receive the echo light signals returned from the optical cable to obtain the loss characteristics along the optical cable path, which is used to determine the location and type of the fault point and the optical power distribution of the optical cable under test. The optical switch module is used to automatically switch between multiple optical cables, enabling a single OTDR module to poll and detect multiple optical cable lines, thereby achieving online monitoring of multiple channels; The photoelectric conversion module is used to convert the echo optical signal received by the OTDR module into an electrical signal and transmit it to the central processing unit module for subsequent analysis and processing. The communication module is used to transmit the processing results and prediction data to a remote server, monitoring platform or host computer system via wired or wireless means, so as to realize remote data uploading, instruction receiving and remote control functions; The central processing unit module is used for high-speed real-time data processing, feature extraction and fault identification of the electrical signals converted by the photoelectric conversion module. At the same time, it integrates the IALA-TCN-BiLSTM prediction model to realize intelligent prediction of the trend of optical power change. The implementation steps of the prediction method are as follows: Step S1: Use the OTDR module and optical switch module to switch the optical path to collect data on the optical power of each optical cable in the distribution network; Step S2: Preprocess the collected optical power data of each optical cable in the distribution network to obtain preprocessed optical power time series data; Step S3: Use TCN network to extract features from optical power time series data, obtain multi-scale time-dependent features by stacking causal dilated convolution, introduce multi-layer residual block structure, and enhance model stability by residual skip connection and Dropout regularization after the output of each dilated convolution layer. Step S4: Construct an IALA-TCN-BiLSTM fusion prediction model. The optical power temporal features extracted by the TCN network are used as input, and the long-short-term dependencies of the temporal features are modeled by BiLSTM. The output of the fusion prediction model is iteratively optimized by the error between the loss function and the true value, and the hyperparameters of the fusion prediction model are optimized by the IALA algorithm to achieve training of the fusion prediction model. In step S4, the specific process of optimizing the hyperparameters of the fusion prediction model using the IALA algorithm is as follows: Step A: Introduce an averaging search mechanism and set the hyperparameter set and the objective function for optimization; Step B: Randomly generate the initial population; iteratively update IALA, calculating the energy factor at the beginning of each iteration. Maintaining a balance between exploration and development; Step C: In the exploration phase, when For long-distance migration, combined with an averaging search mechanism to enhance global diversity, the formula is as follows: ； In the formula, For the first Only individual lemmings in the iteration step The position vector, For the first The historical best position in the next iteration. As the direction control factor, For Brownian motion random vectors, For a mixed weight vector, For the first Average position in each dimension; When energy factor At this time, i.e., the digging behavior, we introduce truncated Levy flight to improve local search efficiency. The formula for truncated Levy flight is as follows: ； In the formula, The scale parameter controls the distribution shape of the Lévy step size. This is the original Lévy distribution. The dynamic cutoff coefficient is defined by the cutoff boundary condition: T max For the maximum iteration step size, As the initial threshold, It is a symbolic function; The position update of an individual uses a truncated Lévy step size, as shown in the following formula: ； In the formula, For the first The historical best position in the next iteration. For the first In the next iteration, the current individual The position vector; This is the dynamic truncation coefficient; Step D: Perform population update, i.e., when the individual parameters exceed the preset search range after the update. When this happens, a reflection correction strategy is used to pull the solution back into the feasible region, saving the current optimal solution, i.e., the optimal parameter combination output; where, Indicates the first The lower bound of dimensionality Indicates the first The upper bound of a dimension; Step S5: The trained IALA-TCN-BiLSTM fusion prediction model predicts the real-time acquired optical power data. The fusion prediction model sequentially completes feature extraction, sequence modeling, and prediction output to obtain the optical power trend results within the target time window.

2. The prediction method according to claim 1, characterized in that: The prediction system also includes a power supply module, which comprises a voltage regulator circuit and an overcurrent protection circuit. The voltage regulator circuit uses an LM317 three-terminal regulator. The input voltage is 24V DC, and the output voltage is adjusted to 5V DC and 3.3V DC via voltage divider resistors, respectively, to power the central processing unit module and the OTDR module. The overcurrent protection circuit consists of a resettable fuse and a TVS diode connected in parallel. When the current exceeds the threshold, the fuse is triggered to open the circuit, and the TVS diode absorbs transient surge voltage.

3. The prediction method according to claim 1, characterized in that: It also includes real-time acquisition of optical power data of each optical cable in the distribution network. The optical power data of each optical cable is collected in real time by controlling the optical switch. During the acquisition process, the controller switches the optical switch channels sequentially according to the set polling cycle, so that the OTDR module can obtain the current optical power information of each optical cable according to the channel.

4. The prediction method according to claim 1, characterized in that: In step S2, the data preprocessing includes noise reduction, smoothing, and data normalization.

5. The prediction method according to claim 1, characterized in that: In step S3, four residual blocks are constructed, and each layer extracts features of different time spans through causal dilated convolution. After the output of each layer of dilated convolution, the model stability is enhanced by residual skip connections and Dropout regularization. After completing the feature extraction of multi-level residual convolution, the temporal dimension is compressed by global average pooling to generate low-dimensional, high-information-density feature vectors.

6. The prediction method according to claim 1, characterized in that: In step A, the formula for the averaging search mechanism is as follows: ； ； In the formula, Indicates the first Average position in each dimension Indicates the current individual In the Position values ​​in each dimension This represents the randomly selected non-optimal individual r on the th... Position values ​​in each dimension This represents the variance of the average location.

Images