Abnormity monitoring and intelligent early warning method for distributed photovoltaic system

Abstract

The invention belongs to the technical field of artificial intelligence, and relates to a distributed photovoltaic system abnormity monitoring and intelligent early warning method. Key parameters of the distributed photovoltaic system are collected at high frequency, noise filtering and signal normalization processing are carried out on data, then multi-domain feature extraction is carried out, anomaly detection is carried out through auto-encoder reconstruction error analysis, convolutional neural network feature classification and a single-class support vector machine, and an anomaly detection result is output. Establishing a multi-dimensional anomaly evaluation model, and synthesizing obtained evaluation results of all dimensions to obtain a final evaluation result; the method comprises the following steps: carrying out hierarchical classification to identify specific anomaly types, carrying out diagnosis of different types of anomalies, identification of typical anomalies and evaluation of fault severity, outputting the specific anomaly types and severity, and providing detailed diagnosis information for early warning decision making. The system state is comprehensively evaluated, corresponding measures are taken according to the grading early warning mechanism and the abnormity severity degree, and complete distributed photovoltaic system abnormity monitoring and intelligent early warning are formed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a distributed photovoltaic system abnormality monitoring and intelligent early warning method. BACKGROUND

[0002] With the large-scale deployment of distributed photovoltaic systems, system operation safety and stability problems are increasingly prominent. The existing monitoring and early warning technologies mainly have the following deficiencies: the traditional monitoring system based on a centralized architecture usually transmits all data to the cloud for processing, and the data transmission delay is large, which is difficult to meet the demand for rapid response to abnormal states; the existing abnormality detection methods are mostly based on threshold judgment, and lack comprehensive analysis ability for multi-dimensional features under complex working conditions; the common monitoring system focuses on fault diagnosis rather than fault early warning, and it is difficult to realize early identification of abnormal states; the existing systems generally lack self-adaptive learning ability, and are difficult to cope with the dynamic changes of photovoltaic system operation environment and state. SUMMARY

[0003] The purpose of the application is to provide a distributed photovoltaic system abnormality monitoring and intelligent early warning method, which collects key parameters of the distributed photovoltaic system at a high frequency, including but not limited to voltage V, current I, active power P, reactive power Q, temperature T, irradiance Ir, environmental parameter Env and device state parameter of the photovoltaic array, and significantly reduces the system response delay. A multi-dimensional abnormality evaluation model is established, including electrical characteristics, environmental conditions, device state and system performance four-dimensional evaluation models, realizing comprehensive evaluation of system state, combining with a hierarchical early warning mechanism and abnormal severity to take corresponding measures, forming a complete distributed photovoltaic system abnormality monitoring and intelligent early warning.

[0004] The technical scheme of the application is as follows:

[0005] A distributed photovoltaic system abnormality monitoring and intelligent early warning method comprises the following steps:

[0006] S1: Collecting key parameters of the distributed photovoltaic system at a high frequency, including but not limited to voltage V, current I, active power P, reactive power Q, temperature T, irradiance Ir, conversion efficiency η, environmental parameter vector Env and device state parameter θ, forming a state vector S; collecting each parameter in the state vector S according to its corresponding sampling frequency, and outputting a multi-channel time series data matrix X(t);

[0007] S={V,I,P,Q,η,T,I r ,Env,θ};

[0008] S2: The multi-channel time series data matrix X(t) is subjected to noise filtering and signal normalization processing, and then multi-domain feature extraction is performed, including extracting time domain, frequency domain and time-frequency domain features, and outputting a standardized multi-dimensional feature vector F and a time-frequency image feature Igram ;

[0009] S3: Standardize the multi-dimensional feature vector F and the time-frequency image I gram , respectively through auto-encoder reconstruction error analysis, convolutional neural network feature classification, and one-class support vector machine for anomaly detection, and the output results are fused and weighted to obtain the anomaly detection result Decision and the confidence Score anomaly ;

[0010] S4: Establish a multi-dimensional anomaly evaluation model, including electrical characteristics, environmental conditions, equipment status, and system performance evaluation models, input anomaly detection results Decision, multi-dimensional feature vector F, and electrical characteristics, environmental conditions, equipment status, and system performance data, and obtain the final evaluation result by synthesizing the evaluation results of each dimension;

[0011] S5: The final evaluation result obtained by S4 and the multi-dimensional feature vector F obtained by S2 are classified and identified by layer to identify specific abnormal types, and different types of anomalies are diagnosed, typical anomalies are identified, and the severity of the fault is evaluated. The specific abnormal type includes coarse-grained classification C coarse and fine-grained classification C fine , output the specific abnormal type, i.e. fine-grained classification result C fine and severity Severity, to provide detailed diagnostic information for early warning decision.

[0012] The distributed photovoltaic system anomaly monitoring and intelligent early warning method, characterized in that it further comprises the following steps:

[0013] S6: Pre-set warning levels and corresponding thresholds, set up different levels of warning corresponding processing measures, output warning information M fine and response strategy according to the abnormal type C warning and severity Severity of S5, and guide the operation and maintenance personnel to take corresponding measures based on the early warning delivery strategy.

[0014] The distributed photovoltaic system anomaly monitoring and intelligent early warning method, characterized in that it further comprises the following two or one of them:

[0015] Build edge-cloud collaborative computing scheduling strategy, combine the computing tasks and system resource state dynamics of each step, distribute computing tasks to edge or cloud, output optimized computing resource allocation scheme, and ensure efficient operation of the system;

[0016] Continuously collect current running data, perform current model performance evaluation, and update model parameters through online incremental learning to complete continuous optimization of the model and ensure that the detection performance improves continuously with data accumulation.

[0017] Further, in S1, the state vector:

[0018] S = {V, I, P, Q, η, T, I r , Env, θ} (1)

[0019] In the formula, V and I are voltage and current vectors, respectively, P and Q are active power and reactive power, respectively, η

[0020] is the conversion efficiency, T is the temperature vector, I r is the irradiance, Env is the environmental parameter vector, and θ is the device state parameter.

[0021] Different sampling strategies are designed for different parameter characteristics, and the sampling frequency is designed based on the Nyquist theorem:

[0022] f sample ≥ 2·f max ·α (4)

[0023] In the formula, f max is the highest frequency component of the signal, which is usually 5 kHz for voltage and current signals and 50 Hz for power signals; α is the safety factor, which is usually 2-5 to prevent spectral aliasing.

[0024] The sampling frequency of each parameter is set as:

[0025] {f V = f I = 10 kHz, f P = f Q = 1 kHz, f η = 1 kHz, f T = 0.1 Hz, f Ir = 0.1 Hz, f Env = 0.1 Hz, f θ = 1 Hz} (5)

[0026] Voltage and current signals: f max = 5 kHz, sampling frequency 10 kHz = 2 × 5 kHz (satisfies the Nyquist theorem)

[0027] Power signal: f max = 50 Hz, sampling frequency 1 kHz = 20 × 50 Hz (sufficient oversampling)

[0028] Temperature, irradiance, environmental parameters: f max = 0.01 Hz, sampling frequency 0.1 Hz = 10 × 0.01 Hz

[0029] Conversion efficiency, device state parameters: f max = 0.1 Hz, sampling frequency 1 Hz = 10 × 0.1 Hz

[0030] Output multi-channel time series data matrix X(t);

[0031] Collect each parameter in the state vector S according to its corresponding sampling frequency to form a multi-channel time series data matrix X(t):

[0032] X(t) = [V(t1), V(t2),..., V(t n );

[0033] I(t1), I(t2),..., I(t n );

[0034] P(t1), P(t2),..., P(t n );

[0035] Q(t1), Q(t2),..., Q(t n );

[0036] η(t1), η(t2),..., η(t n );

[0037] T(t1), T(t2),..., T(t n );

[0038] Ir(t1), Ir(t2),..., Ir(t n );

[0039] Env(t1), Env(t2),..., Env(t n );

[0040] θ(t1), θ(t2),..., θ(t n )]

[0041] In the formula, X(t) is a 9×N matrix, 9 represents the number of parameter channels, and N is the maximum number of sampling points in the time window, which is determined by the highest sampling frequency:

[0042] Data with different sampling frequencies are unified to the same time reference through interpolation or downsampling to ensure the consistency of the matrix structure.

[0043] Output multi-channel time series data matrix X(t) to provide structured data input for subsequent signal processing.

[0044] Further, in the S2, noise filtering:

[0045] X d =W -1 (T λ (W(X))) (6)

[0046] where X is the original multi-channel time series signal, W represents the wavelet transform operator, and Daubechies wavelet is adopted; T λ is the threshold processing function, and soft threshold processing is adopted; X d is the cleaned signal after denoising, and λ is the threshold parameter, which is adaptively adjusted according to the noise level;

[0047] Signal normalization adopts Z-score standardization:

[0048]

[0049] where μ is the mean vector of each channel signal, and σ is the standard deviation vector of each channel signal; X norm is the standardized signal matrix;

[0050] Then, multi-domain feature extraction is performed, including extraction of time domain F time , frequency domain F freq and time-frequency domain features F time-freq :

[0051] F = {F time , F freq , F time-freq} (8)

[0052] The time domain features include statistical features and morphological features:

[0053] F time = {mean, std, kurtosis, skewness, peak, rms, crest,...} (9)

[0054] where mean is the signal mean, std is the standard deviation, kurtosis is the kurtosis, skewness is the skewness, peak is the peak value, rms is the root mean square value, and crest is the peak factor;

[0055] The frequency domain features are obtained through fast Fourier transform:

[0056] F freq = {f c , bandwidth, harmonic r ratio, energy b and,...} (10)

[0057] where fc is the center frequency, bandwidth is the frequency bandwidth, harmonic r ratio is the harmonic ratio, and energy b and is the frequency band energy;

[0058] Time-frequency domain features are extracted by wavelet packet transform:

[0059] F time-freq = {energy i , entropy i ,...,i = 1,2,...,n} (11)

[0060] In the formula, energy i is the energy of the i-th frequency sub-band; entropy i is the entropy value of the i-th frequency sub-band, i = 1,2,...,n, and n is the number of sub-bands.

[0061] Image features are obtained by converting time series data into a two-dimensional representation:

[0062] I gram = f transform (X seg ) (12)

[0063] In the formula, X seg is a data segment within a time window, and f transform is a signal-to-image transformation function, such as a time-frequency map or a recurrence plot.

[0064] The output is a normalized multi-dimensional feature vector F and a time-frequency image I gram .

[0065] The construction process of the feature vector F:

[0066] F = concatenate (F time , F freq , F time-freq )

[0067] Where: F time is calculated by formula (9) and contains 7-dimensional statistical features; F freq is calculated by formula (10) and contains 4-dimensional frequency domain features; Ftime-freq is calculated by formula (11) and contains n-dimensional time-frequency features; concatenate represents a feature concatenation operation.

[0068] The final multi-dimensional feature vector F has a dimension of (7+4+n) x 1, which provides input for subsequent anomaly detection.

[0069] Further, in S3,

[0070] The autoencoder reconstruction error analysis:

[0071] h = f encoder (X; θ e ) (13)

[0072]

[0073] where f encoder is the encoder function, X is the normalized multi-dimensional feature vector F as input; h is the low-dimensional hidden layer representation vector output by the encoder; is the reconstructed feature vector by the decoder; θ e and θ d are the encoder and decoder parameters, including weights and biases; L reconstruction is the reconstruction loss, measuring the difference between the input and the reconstructed output;

[0074] The anomaly score is calculated based on the reconstruction error:

[0075]

[0076] where Score AE (X) is the autoencoder anomaly score, the larger the value, the more abnormal;

[0077] The convolutional neural network feature classification:

[0078] h conv = CNN(I gram ; θ cnn ) (17)

[0079] p(y | I gram ) = softmax(W · h conv + b) (18)

[0080] where I gram is the input image, with size H × W × C; h conv is the convolutional feature, obtained by multi-layer convolution extraction; θ cnn is the set of all trainable parameters of the CNN model, including convolution kernel weights and biases;

[0081] p(y | I gram ) is the class posterior probability distribution given the input image; W is the classifier weight matrix; b is the classifier bias vector;

[0082] The CNN model adopts a lightweight architecture, suitable for edge device deployment:

[0083] CNN = Conv(3, 16, 3, 1) - BN - ReLU - MP(2) - Conv(16, 32, 3, 1) - BN - ReLU - MP(2) - FC(128) - FC(n class ) (19)

[0084] In the formula, Conv(in, out, k, s) represents a convolution layer, and the parameters are input channel number, output channel number, convolution kernel size and step size in turn, BN represents a batch normalization layer, ReLU represents an activation function, MP(k) represents a maximum pooling layer, and the parameter is a pooling kernel size, FC(n) represents a fully connected layer, and the parameter is a neuron number;

[0085] The single-class support vector machine performs anomaly detection:

[0086]

[0087] In the formula, K(x, x i ) is a kernel function, α i is a Lagrange multiplier, ρ is a bias term, and N s is a support vector number;

[0088] The output result is fused and weighted to obtain an anomaly detection result Decision and a confidence score Score anomaly :

[0089] The output process is described as follows:

[0090] 1. Three detection models output: Score AE (X) (formula 16), p(y|I gram ) (formula 18), and f OCSVM (x) (formula 20) respectively;

[0091] 2. The outputs of the three models are fused and weighted according to formula (21) to obtain a comprehensive anomaly score Score anomaly .

[0092] 3. Score anomaly is compared with an adaptive threshold value, and a binary decision result Decision is generated according to formula (22);

[0093] 4. The value of the confidence score Score anomaly is output simultaneously, which provides a quantitative basis for subsequent evaluation;

[0094] The real-time anomaly detection result Decision and the confidence score Score anomaly are output, which provide a basic judgment for subsequent multi-dimensional evaluation;

[0095] Score anomaly = w1·Score AE + w2·(1-max(p(y = normal|I gram )))+w3·D OCSVM (21)

[0096] where w1, w2, w3 are weight coefficients satisfying w1+w2+w3=1 according to the performance of each model; Score AE is the reconstruction error score of the autoencoder; max(p(y=normal|I gram ) is the maximum prediction probability of the CNN model for the normal class; D OCSVM is the distance of the sample to the decision boundary of the OCSVM;

[0097] The abnormality detection result is obtained by comparing the comprehensive score with the adaptive threshold:

[0098]

[0099] where Decision is the final abnormality detection decision result, taking values {Normal, Anomaly};

[0100] Threshold adaptive is the adaptive threshold, which is dynamically adjusted according to the distribution of historical data;

[0101] If the comprehensive abnormality score Score anomaly is less than or equal to the adaptive threshold, it is determined to be normal, otherwise it is determined to be abnormal, and the output Decision is the final abnormality detection result.

[0102] The adaptive threshold is dynamically adjusted based on the distribution of historical data:

[0103] Threshold adaptive = μ Score + β·σ Score (23)

[0104] where μ Score and σ Score are the mean and standard deviation of the score respectively, and β is the adjustment coefficient;

[0105] The output real-time abnormality detection result Decision and the confidence score Score anomaly provide a basis for subsequent multi-dimensional evaluation.

[0106] Further, S4 specifically refers to,

[0107] establishing a multi-dimensional abnormality evaluation model, including an electrical characteristic evaluation model E electrical , an environmental condition evaluation model E environmental , a device state evaluation model E device , and a system performance evaluation model E performance ;

[0108] E comprehensive ={E electrical ,E environmental ,Edevice ,E performance} (24)

[0109] Electrical characteristics evaluation model

[0110] The electrical characteristics evaluation focuses on and is not limited to the abnormality of electrical parameters such as voltage, current, power, etc.:

[0111] E electrical = f e (V, I, P, Q, PF, THD,...) (25)

[0112] In the formula, V, I, P, and Q are voltage, current, active power, and reactive power, respectively, f e is an electrical characteristics evaluation function based on expert rules and data-driven methods; PF is a power factor, and THD is a harmonic distortion rate;

[0113] Environmental condition evaluation model

[0114] The environmental condition evaluation considers the influence of temperature, humidity, and irradiance environmental factors on the photovoltaic system:

[0115] E environmental = f env (T, H, I r , W s , Dust,...) (26)

[0116] In the formula, f env is an environmental condition evaluation function, T, H, I r , W s , and Dust are temperature, humidity, irradiance, wind speed, and dust deposition, respectively;

[0117] Device state evaluation model

[0118] The device state evaluation focuses on the health condition of the components and the inverter device:

[0119] E device = f d (Aging, Hot spot , Connection, Inverter status ,...) (27)

[0120] In the formula, Aging represents the degree of component aging, Hot spot represents the hot spot condition, Connection represents the connection state, and Inverter status represents the inverter state;

[0121] System performance evaluation model

[0122] System performance evaluation focuses on the overall power generation performance of the photovoltaic system:

[0123] E performance = f p (η, PR, Yield,...) (28)

[0124] where η is the conversion efficiency, PR is the performance ratio, and Yield is the power generation;

[0125] Fuzzy logic comprehensive evaluation

[0126] The evaluation results of each dimension are obtained through a fuzzy logic system to obtain the final evaluation:

[0127]

[0128] where A1, A2, etc. are fuzzy sets of input variables, and B i is the fuzzy set of output variables;

[0129] The fuzzy reasoning process is based on the Mamdani reasoning model:

[0130]

[0131] where represents the membership of x j to the fuzzy set A j , and represents the membership of the output y to the fuzzy set B i ;

[0132] The final evaluation result is obtained through centrifugation:

[0133]

[0134] where y i is the output sampling point, and μ(y i ) is its membership;

[0135] The output four-dimensional evaluation vector E comprehensive and the comprehensive health index E final provide detailed state information for abnormal type identification;

[0136] Ecomprehensive = [Eelectrical, Eenvironmental, Edevice, Eperformance]

[0137] where each dimension evaluation result is processed through normalization:

[0138]

[0139] The final comprehensive health index Efinal By fuzzy logic inference of formula (31), the value range is [0, 1], 0 represents a serious anomaly, and 1 represents complete normality.

[0140] Further, S5 refers to:

[0141] Input data: four-dimensional evaluation vector E output by S4 comprehensive and comprehensive health degree index E final , and standardized multi-dimensional feature vector F output by S2

[0142] Data fusion: F enhanced = concatenate (F, E comprehensive , E final )

[0143] Hierarchical classification: coarse-grained and fine-grained classification using F enhanced

[0144] Output: anomaly type C fine and severity Severity

[0145] S51: hierarchical classification identification

[0146] Coarse-grained classification uses global features:

[0147] F global = F enhanced

[0148] C coarse = f classifier1 (F global ) (32)

[0149] In the formula, F global is a global feature vector containing all dimensional features, equal to the enhanced feature vector F enhanced ; F enhanced fuses the original features F, evaluation results E comprehensive and enhanced features of health degree E final ; f classifier1 is a coarse-grained classifier function used to distinguish abnormal categories; C coarse is the coarse-grained classification result, taking values {electrical anomaly, environmental anomaly, device anomaly, performance anomaly}.

[0150] Fine-grained classification uses specific features:

[0151] F specific = feature selection (F enhanced , C coarse )

[0152] C fine = f​classifier2 (F specific |C coarse ) (33)

[0153] where F specific is the feature subset selected for the specific category according to the coarse-grained classification result C coarse ; feature selection is the feature selection function to select relevant features according to the coarse-grained result; f classifier2 is the fine-grained classifier function optimized for the specific abnormal type C fine is the fine-grained classification result, specific abnormal type such as {hot spot fault, component aging, shadow occlusion, etc.};

[0154] S52: Diagnosis of different types of abnormalities

[0155] For different types of abnormalities, design a diagnostic model:

[0156]

[0157] where, is the i-th abnormality-specific diagnostic function; X i is the feature subset related to abnormal type i, Θ i is the diagnostic model parameter; g i is the feature transformation function for the i-th abnormality;

[0158] S53: Identification of typical abnormalities

[0159] Common abnormal types and their feature manifestations:

[0160] Hot spot pattern:

[0161] Pattern hotspot = {T local >T threshold , ΔT> ΔT threshold , I string <I normal ,...} (35)

[0162] where T local is the local temperature measurement value; T threshold is the hot spot detection temperature threshold, usually 80°C; I string is the string current value; I norma is the string current value under normal working conditions;

[0163] Performance degradation:

[0164] Pattern degradation = {P actual / P rated < ηthreshold Δη / Δt > slope threshold ,...} (36)

[0165] where P actual is the actual power output (W); P rated is the rated power output (W); Δη / Δt is the power decay slope; slope threshold is the performance decay threshold;

[0166] Shadowing:

[0167] Pattern shading = {I string <I normal , P < P expected when I r >I r,threshold , I mismatch ,...} (37)

[0168] where I string is the affected group string current, I normal is the normal group string current (A), P expected is the irradiance-based expected power, I r is the current irradiance, I r,threshold is the current difference threshold, I mismatch is the mismatch current indicator;

[0169] Inverter failure:

[0170] Pattern inverterfault = {V output abnormal, THD > THD threshold , efficiency < η invt,hreshold ,...} (38)

[0171] where V output is the inverter output voltage (V), THD hreshold is the voltage harmonic distortion threshold, efficiency is the inverter conversion efficiency, η inv,threshold is the efficiency threshold;

[0172] S54: Failure severity assessment

[0173] Severity = α1 · D feature + α2 · T duration + α3 · I impact (39)

[0174] where D feature is the degree of deviation of the feature from the normal value, T durationFor anomaly duration, I impact For anomaly impact on system performance, α1, α2, α3 are weight coefficients;

[0175] Output specific anomaly type (fine-grained classification result) C fine And severity Severity, provide detailed diagnostic information for early warning decision;

[0176] Severity evaluation integration:

[0177] Severity = λ1·D feature + λ2·T duration + λ3I impact

[0178] Where: λ1, λ2, λ3 are weight coefficients, corresponding to the importance of feature deviation, duration, impact; D feature The degree of deviation from the normal value of the feature is based on F enhanced Calculation; T duration The duration of the anomaly is calculated from the time series data X(t); I impact The degree of impact of the anomaly on the system performance is quantified in combination with Efinal health index; Severity is the fault severity score, with a value range of [0, 1];

[0179] Output unified format:

[0180] The final output is structured data;

[0181] Output S5 ={

[0182] anomaly type : C fine ,

[0183] severity score : Severity,

[0184] confidence: max(p(C fine | F enhanced )),

[0185] feature importance : ranking(F enhanced ),

[0186] temporal info : T duration

[0187] }

[0188] This ensures that the output C fineAnd Severity is consistent with the output of the previous steps in data flow and logic, forming a complete abnormal diagnosis chain.

[0189] Further, S6 refers to:

[0190] S61: Set the pre-warning level and set the corresponding threshold

[0191] The hierarchical warning mechanism divides the system state into five levels: normal, attention, warning, alarm, and danger, according to the abnormal evaluation results:

[0192]

[0193] In the formula, L1, L2, L3, L4 are pre-warning thresholds; E final is the comprehensive health index, taking values [0, 1]; Level is the warning level, taking values {Normal, Attention, Warning, Alarm, Danger};

[0194] S62: Set the corresponding processing measures for different levels of warning

[0195] Define the warning strategy matrix

[0196] Strategy(Level, Type) = Action specific (41)

[0197] In the formula: Strategy is the warning strategy matrix, defining the corresponding processing measures for different levels and types, Level is the warning level, Type is the abnormal type, Action specific is the targeted response action;

[0198] S63: Warning information generation

[0199] Warning information generation model:

[0200] M warning = {Level, Type, Location, Time, Severity, Suggestion, Confidence} (42)

[0201] In the formula, Level is the warning level, Type is the abnormal type, Location is the abnormal position, Time is the detection time, Severity is the severity, Suggestion is the processing suggestion, and Confidence is the confidence;

[0202] S64: Warning delivery strategy

[0203] The warning delivery priority is based on importance evaluation:

[0204] P transmission =f t (Level,Type,Receiver) (43)

[0205] In the formula, f t For the early warning propagation strategy function, P transmission To prioritize the transmission of early warnings, the Receiver is the recipient;

[0206] Output complete early warning information M warning And response strategies to guide operations and maintenance personnel to take appropriate measures.

[0207] The aforementioned method for anomaly monitoring and intelligent early warning in distributed photovoltaic systems constructs an edge-cloud collaborative computing scheduling strategy. Combining the computing tasks and dynamic system resource status of each step, it allocates computing tasks to the edge or cloud, outputting an optimized computing resource allocation scheme to ensure efficient system operation. The method includes:

[0208] S71: Task Scheduling Strategy

[0209] The edge-cloud collaborative computing architecture adopts a layered design, dynamically allocating tasks based on computing needs and resource conditions:

[0210] Task Allocation =f scheduler (Task Property Resource State (44)

[0211] In the formula, Task Allocation Assign results to tasks and specify the execution location for each task; f scheduler This is a task scheduler function based on a heuristic algorithm; Task Property Including computational complexity, time sensitivity, and resource State This includes computing power, load status, and resource information;

[0212] S72: Multi-objective optimization

[0213] Task scheduling strategies are based on multi-objective optimization:

[0214] minJ=w1·T processing +w2·E consumption +w3·B usage (45)

[0215] stT processing ≤T deadline (46)

[0216] E consumption ≤E budget(47)

[0217] Resource Usage ≤Resource Capacity (48)

[0218] where T processing is the processing time, E consumption is the energy consumption, B usage is the bandwidth usage, w1, w2, w3 are weight coefficients; T deadline is the task deadline, E budge t is the energy budget (J), Resource Capacity is the resource capacity upper limit;

[0219] S73: Model compression and deployment

[0220] Edge-side lightweight model deployment adopts model compression technology:

[0221] Model edge = f compress (Model original , Compression Ratio ) (49)

[0222] Compression methods include pruning, quantization and knowledge distillation:

[0223] Model pruned = f prune (Model original , Pruning Strategy ) (50)

[0224] Model quantized = f quantize (Model pruned , Bit Width ) (51)

[0225] Model distilled = f distill (Model teacher , Model student ) (52)

[0226] where Model edge is the edge-side lightweight model, f compress is the model compression function, Model original is the original complete model, Compression Ratio is the compression ratio, Modelp runed is the pruned model, f prune is the pruning function, Pruning StrategyFor pruning strategy, Model quantized For quantized model, f quantize For quantization function, Bit Width For quantization bit-width, Model distilled For distilled model, f distill For knowledge distillation function, Model teacher For teacher model, Model student For student model;

[0227] S74: Edge capability evaluation

[0228] Edge device computing capability evaluation model:

[0229] C edge = {CPU: f cpu , RAM: m ram , Storage: s storage , Power: p max} (53)

[0230] In the formula, C edge is the comprehensive computing capability score of the edge device, f cpu is the CPU performance index, m ram is the memory capacity, s storage is the storage capacity (GB), and p max is the maximum power consumption;

[0231] S75: Data interaction strategy

[0232] Edge-cloud data interaction strategy:

[0233] D upload = f filter (D raw , Policy upload ) (54)

[0234] Policy upload = {Rule1, Rule2, …, Rule n} (55)Rule i : IF Condition i THEN Action i (56)

[0235] In the formula, D upload is the filtered data that needs to be uploaded to the cloud, f filter is the data filtering function, D raw is the original collected data, Policy upload is the data upload strategy set, and Rule i: the i-th rule, Condition i : the trigger condition of the i-th rule, Action i : the execution action corresponding to the i-th rule.

[0236] Output the optimized computing resource allocation scheme to ensure efficient operation of the system;

[0237] Continuously collect current operation data, evaluate the current model performance, and update the model parameters through online incremental learning to complete the continuous optimization of the model, ensuring that the detection performance improves continuously with data accumulation, specifically including:

[0238] S81: Incremental learning update

[0239] The online model update mechanism is based on incremental learning technology to achieve continuous optimization of the model:

[0240] Θ t+1 = Θ t + ΔΘ t (57)

[0241] ΔΘ t = f update (D new , Θ t , Performance t ) (58)

[0242] where Θ t is the model parameter at time t, D new is the newly collected data, f update is the parameter update function, and Performance t is the current performance evaluation;

[0243] S82: Incremental learning loss

[0244] Incremental learning strategy design:

[0245] L incremental = L new + λ·L memory (59)

[0246] where L incremental is the total loss function of incremental learning, L new is the loss on new data, L memory is the memory maintenance loss to prevent catastrophic forgetting, and λ is the balance parameter between new and old knowledge;

[0247] S83: Knowledge distillation update

[0248] Knowledge distillation is used for model update:

[0249]

[0250] where L CE is the cross-entropy loss between the true label y and the predicted result KL is the KL divergence, z t and z t-1 are the outputs of the current model and the historical model, respectively, τ is the temperature parameter, and α is the balance coefficient.

[0251] S84: Adaptive threshold adjustment

[0252] Adaptive threshold adjustment strategy:

[0253] Threshold t+1 = Threshold t + γ · (FAR t - Target FAR ) (61)

[0254] where FAR t is the current false positive rate, Target FAR is the target false positive rate, and γ is the adjustment coefficient.

[0255] S85: Performance evaluation index

[0256] Feature importance evaluation and selection:

[0257] Importance(F i ) = Information Gain (Class, F i ) (62)

[0258] IG(Class, F i ) = H(Class) - H(Class | F i ) (63)

[0259] where Importance(F i ) is the importance score of the i-th feature, Information Gain is the information gain, which measures the classification ability of the feature; F i is the i-th feature, Class is the class label, IG(Class, F i ) is the information gain of feature Fi on class Class; H(Class) is the class entropy, and H(Class | F i ) is the class entropy given the feature F i .

[0260] Model performance evaluation index: ​

[0261]

[0262]

[0263] In the formula, TP, FP, and FN represent the number of true positive, false positive, and false negative samples, respectively; Precision is the accuracy rate, the proportion of samples correctly predicted as abnormal; Recall is the recall rate, the proportion of actual abnormalities correctly identified; and F 1Score is the harmonic mean of precision and recall; AUROC is the area under the ROC curve, which measures the overall performance of the classifier; t is the classification threshold parameter; TPR is the true positive rate, and FPR is the false positive rate.

[0264] The present invention has the following advantages:

[0265] 1. Reduce the response latency of the anomaly monitoring system to achieve real-time detection of abnormal states;

[0266] 2. Improve the accuracy of anomaly detection and reduce false alarms and false negatives;

[0267] 3. Enable early identification of fault precursors and provide sufficient processing time window;

[0268] 4. Efficiently deploy complex detection algorithms on resource-constrained edge devices;

[0269] 5. Construct a scientific and reasonable anomaly assessment and early warning mechanism to achieve tiered response. Detailed Implementation

[0270] A method for anomaly monitoring and intelligent early warning in distributed photovoltaic systems includes the following steps:

[0271] S1: High-frequency acquisition of key parameters of the distributed photovoltaic system, including but not limited to the photovoltaic array voltage V, current I, active power P, reactive power Q, temperature T, irradiance Ir, η is the conversion efficiency, environmental parameter vector Env and equipment state parameter θ, forming a state vector S; each parameter in the state vector S is acquired according to its corresponding sampling frequency, and a multi-channel time series data matrix X(t) is output.

[0272] S={V,I,P,Q,η,T,I r ,Env,θ};

[0273] S2: The multi-channel time-series data matrix X(t) undergoes noise filtering and signal normalization, followed by multi-domain feature extraction, including extraction of time-domain, frequency-domain, and time-frequency-domain features, outputting a standardized multi-dimensional feature vector F and time-frequency image features I. gram ;

[0274] S3: Standardize the multi-dimensional feature vector F and the time-frequency image I gram , respectively through auto-encoder reconstruction error analysis, convolutional neural network feature classification, and one-class support vector machine for anomaly detection, and the output results are fused and weighted to obtain the anomaly detection result Decision and the confidence Score anomaly ;

[0275] S4: Establish a multi-dimensional anomaly evaluation model, including electrical characteristics, environmental conditions, equipment status, and system performance four-dimensional evaluation models, input anomaly detection results Decision, multi-dimensional feature vector F, and electrical characteristics, environmental conditions, equipment status, and system performance data, and obtain the final evaluation result by synthesizing the evaluation results of each dimension;

[0276] S5: The final evaluation result obtained in S4 and the multi-dimensional feature vector F obtained in S2 are classified and identified in layers to identify specific abnormal types, and diagnose different types of abnormalities, typical abnormalities, and fault severity. The identification of specific abnormal types includes coarse-grained classification C coarse and fine-grained classification C fine , output the specific abnormal type, i.e., fine-grained classification result C fine and severity Severity, to provide detailed diagnostic information for early warning decision-making.

[0277] Also includes:

[0278] S6: Pre-set early warning levels and corresponding thresholds, set up corresponding processing measures for different levels of early warning, output early warning information M fine and response strategy based on the abnormal type C warning and severity Severity of S5, and guide the operation and maintenance personnel to take corresponding measures based on the early warning delivery strategy.

[0279] Also includes: constructing an edge-cloud collaborative computing scheduling strategy, combining the computing tasks and system resource states of each step to dynamically allocate computing tasks to the edge or cloud, outputting an optimized computing resource allocation scheme to ensure efficient operation of the system;

[0280] Continuously collect current operation data, perform current model performance evaluation, and update model parameters through online incremental learning to complete continuous optimization of the model, ensuring that the detection performance continuously improves as data accumulates.

[0281] In S1, the state vector is:

[0282] S = {V, I, P, Q, η, T, I r , Env, θ} (1)

[0283] In the formula, V, I are voltage and current vectors respectively, P, Q are active power and reactive power respectively, η

[0284] η is conversion efficiency, T is temperature vector, I r is irradiance, Env is environmental parameter vector, θ is device state parameter

[0285] Different sampling strategies are designed for different parameter characteristics, and the sampling frequency is designed based on the Nyquist theorem:

[0286] f sample ≥2·f max ·α (4)

[0287] In the formula, f max is the highest frequency component of the signal, usually 5kHz for voltage and current signals and 50Hz for power signals; α is the safety factor, usually 2-5 to prevent spectral aliasing

[0288] The sampling frequency of each parameter is set as:

[0289] {f V =f I =10kHz,f P =f Q =1kHz,f η =1kHz,f T =0.1Hz,f Ir =0.1Hz,f Env =0.1Hz,f θ =1HZ}(5)

[0290] Corresponding to the f max relationship described above:

[0291] Voltage and current signals: f max = 5kHz, sampling frequency 10kHz = 2x5kHz (satisfies the Nyquist theorem)

[0292] Power signals: f max = 50Hz, sampling frequency 1kHz = 20x50Hz (sufficient oversampling)

[0293] Temperature, irradiance, environmental parameters: f max = 0.01Hz, sampling frequency 0.1Hz = 10x0.01Hz

[0294] Conversion efficiency, device state parameters: f max = 0.1Hz, sampling frequency 1Hz = 10x0.1Hz

[0295] Output multi-channel time series data matrix X(t)

[0296] Each parameter in the state vector S is collected according to its corresponding sampling frequency to form multi-channel time series data:

[0297] X(t) = [V(t1), V(t2),..., V(t n );

[0298] I(t1), I(t2),..., I(t n );

[0299] P(t1), P(t2),..., P(t n );

[0300] Q(t1), Q(t2),..., Q(t n );

[0301] η(t1), η(t2),..., η(t n );

[0302] T(t1), T(t2),..., T(t n );

[0303] Ir(t1), Ir(t2),..., Ir(t n );

[0304] Env(t1), Env(t2),..., Env(t n );

[0305] θ(t1), θ(t2),..., θ(t n )]

[0306] In the formula, X(t) is a 9×N matrix, 9 represents the number of parameter channels, and N is the maximum number of sampling points in a time window, which is determined by the highest sampling frequency: where T window is the data collection time window;

[0307] Data with different sampling frequencies are unified to the same time reference through interpolation or down-sampling to ensure the consistency of the matrix structure;

[0308] The multi-channel time series data matrix X(t) is output, providing structured data input for subsequent signal processing.

[0309] In S2, noise filtering:

[0310] X d = W -1 (T λ (W(X))) (6)

[0311] In the formula, X is the original multi-channel time sequence signal, W represents the wavelet transform operator, and the Daubechies wavelet is adopted; T λ is a threshold processing function, and the soft threshold processing is adopted; X d is the cleaned signal after denoising, and λ is a threshold parameter, which is adaptively adjusted according to the noise level;

[0312] The signal normalization adopts Z-score standardization:

[0313]

[0314] In the formula, μ is a mean vector of the channel signals, and σ is a standard deviation vector of the channel signals; X norm is a standardized signal matrix;

[0315] Then, multi-domain feature extraction is performed, including extraction of time domain F time , frequency domain F freq and time-frequency domain features F time-freq :

[0316] F = {F time , F freq , F time-freq} (8)

[0317] The time domain features include statistical features and morphological features:

[0318] F time = {mean, std, kurtosis, skewness, peak, rms, crest,...} (9)

[0319] In the formula: mean is the signal mean, std is the standard deviation, kurtosis is the kurtosis, skewness is the skewness, peak is the peak value, rms is the root mean square value, and crest is the peak factor;

[0320] The frequency domain features are obtained through fast Fourier transform:

[0321] F freq = {f c , bandwidth, harmonic r ratio, energy b and,...} (10)

[0322] In the formula: fc is the center frequency, bandwidth is the frequency bandwidth, harmonic r ratio is the harmonic ratio, and energy b and is the frequency band energy;

[0323] Time-frequency domain features are extracted by wavelet packet transform:

[0324] F time-freq i ,entropy i ,...,i=1,2,...,n} (11)

[0325] energy i is the energy of the i-th frequency sub-band; entropy i is the entropy value of the i-th frequency sub-band, i=1,2,...,n, and n is the number of sub-bands;

[0326] Image features are obtained by converting time series data into a two-dimensional representation:

[0327] I gram =f transform (X seg ) (12)

[0328] where X seg is a data segment within a time window, and f transform is a signal-to-image transformation function, such as a time-frequency map or a recurrence plot;

[0329] The output is a normalized multi-dimensional feature vector F and a time-frequency image I gram ;

[0330] The construction process of the feature vector F:

[0331] F=concatenate(F time ,F freq ,F time-freq )

[0332] where F time is calculated by equation (9) and contains 7-dimensional statistical features; F freq is calculated by equation (10) and contains 4-dimensional frequency domain features; Ftime-freq is calculated by equation (11) and contains n-dimensional time-frequency features;

[0333] concatenate represents a feature concatenation operation.

[0334] The final multi-dimensional feature vector F has a dimension of (7+4+n)×1, providing input for subsequent anomaly detection.

[0335] In S3,

[0336] The autoencoder reconstruction error analysis:

[0337] h=f encoder (X;θ e ) (13)​

[0338]

[0339]

[0340] where f encoder is the encoder function, X is the normalized multi-dimensional feature vector F as input; h is the low-dimensional hidden layer representation vector output by the encoder;

[0341] is the reconstructed feature vector by the decoder; θ e and θ d are the encoder and decoder parameters, including weights and biases; L reconstruction is the reconstruction loss, measuring the difference between the input and the reconstructed output;

[0342] The anomaly score is calculated based on the reconstruction error:

[0343]

[0344] where Score AE (X) is the autoencoder anomaly score, the larger the value, the more abnormal;

[0345] The convolutional neural network feature classification:

[0346] h conv = CNN(I gram ; θ cnn ) (17)

[0347] p(y | I gram ) = softmax(W · h conv + b) (18)

[0348] where I gram is the input image with size H × W × C; h conv is the convolutional feature extracted by multiple layers of convolution; θ cnn is the set of all trainable parameters of the CNN model, including convolution kernel weights and biases;

[0349] p(y | I gram ) is the class posterior probability distribution given the input image; W is the classifier weight matrix; b is the classifier bias vector;

[0350] The CNN model adopts a lightweight architecture, suitable for edge device deployment:

[0351] CNN = Conv(3, 16, 3, 1) - BN - ReLU - MP(2) - Conv(16, 32, 3, 1) - BN - ReLU - MP(2) - FC(128) - FC(n class )(19)

[0352] where Conv(in, out, k, s) represents a convolution layer, the parameters are input channel number, output channel number, kernel size and stride in turn, BN represents a batch normalization layer, ReLU represents an activation function, MP(k) represents a max-pooling layer, the parameter is pooling kernel size, FC(n) represents a fully connected layer, the parameter is neuron number;

[0353] The single-class support vector machine performs anomaly detection:

[0354]

[0355] where K(x, x i ) is a kernel function, α i is a Lagrange multiplier, ρ is a bias term, and N s is the number of support vectors;

[0356] The output result is fused and weighted to obtain an anomaly detection result Decision and a confidence Score anomaly : nn

[0357] The output process is described as follows:

[0358] 1. Three detection models output: Score AE (X) (formula 16), p(y|I gram ) (formula 18), and f OCSVM (x) (formula 20), respectively;

[0359] 2. The outputs of the three models are fused and weighted according to formula (21) to obtain a comprehensive anomaly score Score anomaly ;

[0360] 3. Score anomaly is compared with an adaptive threshold, and a binary decision result Decision is generated according to formula (22);

[0361] 4. The value of the confidence Score anomaly is output at the same time, providing a quantitative basis for subsequent evaluation;

[0362] The real-time anomaly detection result Decision and the confidence Score anomaly are output, providing a basis for subsequent multi-dimensional evaluation;

[0363] Score anomaly = w1·ScoreAE +w2·(1-max(p(y=normal|I gram )))+w3·D OCSVM (21)

[0364] where w1, w2, w3 are weight coefficients satisfying w1+w2+w3=1 according to the performance of each model; Score AE is the reconstruction error score of the autoencoder; max(p(y=normal|I gram ) is the maximum prediction probability of the CNN model for the normal class; D OCSVM is the distance of the sample to the decision boundary of the OCSVM.

[0365] The anomaly detection result is obtained by comparing the comprehensive score with the adaptive threshold:

[0366]

[0367] where Decision is the final anomaly detection decision result, taking values {Normal, Anomaly};

[0368] Threshold adaptive is the adaptive threshold, which is dynamically adjusted according to the distribution of historical data;

[0369] If the comprehensive anomaly score Score anomaly is less than or equal to the adaptive threshold, it is determined to be normal, otherwise it is determined to be abnormal, and the output Decision is the final anomaly detection result;

[0370] The adaptive threshold is dynamically adjusted based on the distribution of historical data:

[0371] Threshold adaptive =μ Score +β·σ Score (23)

[0372] where μ Score and σ Score are the mean and standard deviation of the score respectively, and β is the adjustment coefficient;

[0373] The output real-time anomaly detection result Decision and the confidence score Score anomaly provide a basis for subsequent multi-dimensional evaluation.

[0374] S4 specifically refers to,

[0375] A multi-dimensional anomaly evaluation model is established, including an electrical characteristic evaluation model E electrical , an environmental condition evaluation model E environmental , and a device state evaluation model E device, system performance evaluation model E performance ;

[0376] E comprehensive = {E electrical , E environmental , E device , E performance} (24)

[0377] electrical characteristics evaluation model

[0378] The electrical characteristics evaluation focuses on and is not limited to the abnormality of electrical parameters such as voltage, current, power, etc.:

[0379] E electrical = f e (V, I, P, Q, PF, THD,...) (25)

[0380] In the formula, V, I, P, and Q are voltage, current, active power, and reactive power, respectively, f e is an electrical characteristics evaluation function based on expert rules and data-driven methods; PF is a power factor, and THD is a harmonic distortion rate;

[0381] environmental condition evaluation model

[0382] The environmental condition evaluation considers the influence of temperature, humidity, irradiance environmental factors on the photovoltaic system:

[0383] E environmental = f env (T, H, I r , W s , Dust,...) (26)

[0384] In the formula, f env is an environmental condition evaluation function, T, H, I r , W s , and Dust are temperature, humidity, irradiance, wind speed, and dust deposition, respectively;

[0385] equipment state evaluation model

[0386] The equipment state evaluation focuses on the health condition of components and inverter equipment:

[0387] E device = f d (Aging, Hot spot , Connection, Inverter status ,...) (27)

[0388] In the formula, Aging represents the aging degree of components, Hot spotHot spot, Connection, Inverter status Hot spot, Connection, Inverter

[0389] System performance evaluation model

[0390] System performance evaluation focuses on the overall power generation performance of the photovoltaic system:

[0391] E performance = f p (η, PR, Yield,...) (28)

[0392] where η is the conversion efficiency, PR is the performance ratio, and Yield is the power generation;

[0393] Fuzzy logic comprehensive evaluation

[0394] The evaluation results of each dimension are obtained through a fuzzy logic system to obtain the final evaluation:

[0395]

[0396] where A1, A2, etc. are fuzzy sets of input variables, and B i is the fuzzy set of output variables;

[0397] The fuzzy reasoning process is based on the Mamdani reasoning model:

[0398]

[0399] where represents the membership degree of x j to the fuzzy set A j , and the membership degree of the output y to the fuzzy set B i ;

[0400] The final evaluation result is obtained through centrifugation:

[0401]

[0402] where y i is the output sampling point, and μ(y i ) is its membership degree;

[0403] The output four-dimensional evaluation vector E comprehensive and the comprehensive health degree index E final provide detailed state information for abnormal type identification;

[0404] Ecomprehensive = [Eelectrical, Eenvironmental, Edevice, Eperformance]

[0405] where each dimension assessment result is normalized:

[0406]

[0407] Final comprehensive health index E final Obtained by fuzzy logic inference of formula (31), value range [0,1], 0 represents severe abnormality, and 1 represents complete normality.

[0408] S5 refers to:

[0409] Input data: four-dimensional evaluation vector E output by S4 comprehensive and comprehensive health index E final , and standardized multi-dimensional feature vector F output by S2

[0410] Data fusion: F enhanced = concatenate(F,E comprehensive ,E final )

[0411] Hierarchical classification: use F enhanced to perform coarse-grained and fine-grained classification

[0412] Output: abnormal type C fine and severity Severity

[0413] S51: hierarchical classification identification

[0414] Coarse-grained classification uses global features:

[0415] F global = F enhanced

[0416] C coarse = f classifier1 (F global ) (32)

[0417] In the formula, F global is a global feature vector containing all dimension features, equal to enhanced feature vector F enhanced ; F enhanced fuses original features F, enhanced features of evaluation results E comprehensive and health E final ; f classifier1 is a coarse-grained classifier function used to distinguish abnormal categories; C coarse is a coarse-grained classification result, taking values {electrical abnormality, environmental abnormality, equipment abnormality, performance abnormality}.

[0418] Fine-grained classification uses specific features:

[0419] Fspecific =feature selection (F enhanced C coarse )

[0420] C fine =f classifier2 (F specific |C coarse (33)

[0421] In the formula, F specific Based on the coarse-grained classification result C coarse Select a specific subset of category-specific features; selection The feature selection function is used to select relevant features based on the coarse-grained results; f classifier2 For fine-grained classifier functions, optimize C for specific anomaly types. fine For fine-grained classification results, specific anomaly types include {hot spot failure, component aging, shadow occlusion, etc.};

[0422] S52: Diagnosis of different types of abnormalities

[0423] Design diagnostic models for different types of anomalies:

[0424]

[0425] In the formula, X is a dedicated diagnostic function for the i-th type of anomaly; i For the feature subset related to anomaly type i, Θ i For diagnostic model parameters; g i Let i be the feature transformation function for the i-th type of anomaly;

[0426] S53: Identification of Typical Anomalies

[0427] Common anomaly types and their characteristic manifestations:

[0428] Hot spot mode:

[0429] Pattern hotspot ={T local >T threshold ,ΔT>ΔT threshold ,I string <I normal ,...} (35)

[0430] In the formula, T local This is a local temperature measurement value; T threshold The temperature threshold for hot spot detection is typically 80°C; I string I represents the string current value. norma l represents the string current value under normal operating conditions;

[0431] Performance degradation:

[0432] Pattern degradation = {P actual / P rated < η threshold , Δη / Δt > slope threshold ,...} (36)

[0433] where P actual is the actual power output (W); P rated is the rated power output (W); Δη / Δt is the power degradation slope; slope threshold is the performance degradation determination threshold.

[0434] Shadowing:

[0435] Pattern shading = {I string < I normal , P < P expected when I r > I r,threshold , I mismatch ,...} (37)

[0436] where I string is the affected group string current, I normal is the normal group string current (A), P expected is the expected power based on irradiance, I r is the current irradiance, I r,threshold is the current difference threshold, I mismatch is the mismatch current indicator;

[0437] Inverter failure:

[0438]

[0439] where V output is the inverter output voltage (V), THDt hreshold is the voltage harmonic distortion threshold, efficiency is the inverter conversion efficiency, η inv,threshold is the efficiency threshold;

[0440] S54: Failure severity assessment

[0441] Severity = α1·D feature + α2·T duration + α3·I impact (39)

[0442] where D featureT represents the degree to which a feature deviates from its normal value. duration For the duration of the abnormality, I impact The impact of anomalies on system performance is represented by α1, α2, and α3, which are weighting coefficients.

[0443] Output the specific anomaly type (fine-grained classification result) C fine The severity level provides detailed diagnostic information for early warning decisions.

[0444] Severity assessment integration:

[0445] Severity=λ1·D feature +λ2·T duration +λ3I impact

[0446] Where: λ1, λ2, and λ3 are weight coefficients, corresponding to the importance of feature deviation, duration, and degree of influence, respectively; D feature To indicate the degree to which a feature deviates from its normal value, based on F enhanced Calculation; T duration The duration of the anomaly is statistically derived from the time-series data X(t); I impact The impact of the anomaly on system performance is quantified by combining the Efinal health index; Severity is the score for the severity of the fault, with a value range of [0,1].

[0447] Output in a uniform format:

[0448] The final output is structured data;

[0449] Output S5 ={

[0450] anomaly type :C fine ,

[0451] severity score Severity

[0452] confidence: max(p(C) fine |F enhanced )),

[0453] feature importance :ranking(F enhanced ),

[0454] temporal info :T duration

[0455] }

[0456] This ensures that the output C of S5 fine And Severity is consistent with the output of the previous steps in data flow and logic, forming a complete chain of abnormal diagnosis.

[0457] S6 refers to:

[0458] S61: Set the pre-warning level and set the corresponding threshold

[0459] The hierarchical warning mechanism divides the system state into five levels: normal, attention, warning, alarm, and danger, according to the abnormal evaluation results:

[0460]

[0461] In the formula, L1, L2, L3, L4 are pre-warning thresholds; E final is the comprehensive health index, taking values [0, 1]; Level is the pre-warning level, taking values {Normal, Attention, Warning, Alarm, Danger};

[0462] S62: Set the corresponding processing measures for different levels of pre-warning

[0463] Define the pre-warning strategy matrix

[0464] Strategy(Level, Type) = Action specific (41)

[0465] In the formula: Strategy is the pre-warning strategy matrix, defining the corresponding processing measures for different levels and types, Level is the pre-warning level, Type is the abnormal type, Action specific is the targeted response action;

[0466] S63: Pre-warning information generation

[0467] Pre-warning information generation model:

[0468] M warning = {Level, Type, Location, Time, Severity, Suggestion, Confidence} (42)

[0469] In the formula, Level is the pre-warning level, Type is the abnormal type, Location is the abnormal position, Time is the detection time, Severity is the severity, Suggestion is the processing suggestion, and Confidence is the confidence;

[0470] S64: Pre-warning delivery strategy

[0471] The early warning delivery priority is based on importance evaluation:

[0472] P transmission = f t (Level, Type, Receiver) (43)

[0473] wherein f t is an early warning delivery strategy function, P transmission is an early warning delivery priority, and Receiver is a receiver;

[0474] Output complete early warning information M warning and response strategies to guide operation and maintenance personnel to take corresponding measures.

[0475] An edge-cloud collaborative computing scheduling strategy is constructed, and the computing tasks and system resource states of each step are combined to dynamically allocate computing tasks to the edge or cloud, output an optimized computing resource allocation scheme, and ensure efficient operation of the system, including:

[0476] S71: Task scheduling strategy

[0477] The edge-cloud collaborative computing architecture adopts a hierarchical design, and dynamically allocates tasks according to computing requirements and resource conditions:

[0478] Task Allocation = f scheduler (Task Property , Resource State ) (44)

[0479] wherein Task Allocation is a task allocation result, specifying the execution location of each task; f scheduler is a task scheduler function based on heuristic algorithms; Task Property includes computing complexity and time sensitivity characteristics, and Resource State includes computing capability and load state resource information;

[0480] S72: Multi-objective optimization

[0481] The task scheduling strategy is based on multi-objective optimization:

[0482] min J = w1·T processing + w2·E consumption + w3·B usage (45)

[0483] s.t. T processing ≤ T deadline (46)

[0484] E consumption ≤ Ebudget (47)

[0485] Resource Usage ≤Resource Capacity (48)

[0486] where T processing is the processing time, E consumption is the energy consumption, B usage is the bandwidth usage, w1, w2, w3 are weight coefficients; T deadline is the task deadline, E budge t is the energy budget (J), Resource Capacity is the resource capacity upper limit;

[0487] S73: Model compression and deployment

[0488] The edge-side lightweight model deployment adopts model compression technology:

[0489] Model edge = f compress (Model original , Compression Ratio ) (49)

[0490] The compression method includes pruning, quantization and knowledge distillation:

[0491] Model pruned = f prune (Model original , Pruning Strategy ) (50)

[0492] Model quantized = f quantize (Model pruned , Bit Width ) (51)

[0493] Model distilled = f distill (Model teacher , Model student ) (52)

[0494] where Model edge is the edge-side lightweight model, f compress is the model compression function, Model original is the original complete model, Compression Ratio is the compression ratio, Modelp runed is the pruned model, f prune is the pruning function, PruningStrategy Model quantized f quantize Bit Width Model distilled f distill Model teacher Model student Model

[0495] S74: Edge capability evaluation

[0496] Edge device computing capability evaluation model:

[0497] C edge = {CPU: f cpu , RAM: m ram , Storage: s storage , Power: p max} (53)

[0498] In the formula, C edge is the comprehensive computing capability score of the edge device, f cpu is the CPU performance index, m ram is the memory capacity, s storage is the storage capacity (GB), and p max is the maximum power consumption.

[0499] S75: Data interaction strategy

[0500] Edge-cloud data interaction strategy:

[0501] D upload = f filter (D raw , Policy upload ) (54)

[0502] Policy upload = {Rule1, Rule2, …, Rule n} (55) Rule i : IF Condition i THEN Action i (56)

[0503] In the formula, D upload is the filtered data that needs to be uploaded to the cloud, f filter is the data filtering function, D raw is the original collected data, Policy upload is the data upload strategy set, and Rule i: the i-th rule, Condition i : the trigger condition of the i-th rule, Action i : the execution action corresponding to the i-th rule.

[0504] Output the optimized computing resource allocation scheme to ensure efficient operation of the system;

[0505] Continuously collect current operation data, evaluate the current model performance, update the model parameters through online incremental learning, and complete the continuous optimization of the model to ensure that the detection performance improves continuously with data accumulation, specifically including:

[0506] S81: Incremental learning update

[0507] The online model update mechanism is based on incremental learning technology to realize the continuous optimization of the model:

[0508] Θ t+1 = Θ t + ΔΘ t (57)

[0509] ΔΘ t = f update (D new , Θ t , Performance t ) (58)

[0510] In the formula, Θ t is the model parameter at time t, D new is the newly collected data, f update is the parameter update function, and Performance t is the current performance evaluation;

[0511] S82: Incremental learning loss

[0512] Incremental learning strategy design:

[0513] L incremental = L new + λ·L memory (59)

[0514] In the formula, L incremental is the total loss function of incremental learning, L new is the loss on new data, L memory is the memory maintenance loss to prevent catastrophic forgetting, and λ is the balance parameter of new and old knowledge;

[0515] S83: Knowledge distillation update

[0516] Knowledge distillation is used for model update:

[0517]

[0518] where L CE is the cross-entropy loss between the true label y and the predicted result KL is the KL divergence, z t and z t-1 are the outputs of the current model and the historical model, respectively, τ is the temperature parameter, and α is the balance coefficient.

[0519] S84: Adaptive threshold adjustment

[0520] Adaptive threshold adjustment strategy:

[0521] Threshold t+1 = Threshold t + γ · (FAR t - Target FAR ) (61)

[0522] where FAR t is the current false positive rate, Target FAR is the target false positive rate, and γ is the adjustment coefficient.

[0523] S85: Performance evaluation index

[0524] Feature importance evaluation and selection:

[0525] Importance(F i ) = Information Gain (Class, F i ) (62)

[0526] IG(Class, F i ) = H(Class) - H(Class | F i ) (63)

[0527] where Importance(F i ) is the importance score of the i-th feature, Information Gain is the information gain, which measures the classification ability of the feature; F i is the i-th feature, Class is the class label, IG(Class, F i ) is the information gain of feature Fi on class Class; H(Class) is the class entropy, and H(Class | F i ) is the class entropy given the feature F i .

[0528] Model performance evaluation index: ​

[0529]

[0530]

[0531] In the formula, TP, FP and FN are the numbers of true positive, false positive and false negative samples respectively; Precision is the precision rate, the proportion of correctly predicted anomalies; Recall is the recall rate, the proportion of actual anomalies correctly identified; F 1Score is the harmonic mean of precision and recall; AUROC is the area under the ROC curve, which measures the overall performance of the classifier; t is the classification threshold parameter; TPR is the true positive rate, and FPR is the false positive rate.

[0532] System overall architecture

[0533] A three-layer architecture is adopted, including a perception layer, an edge layer and a cloud layer. The perception layer is responsible for data acquisition, the edge layer performs real-time anomaly detection, and the cloud layer performs model training and deep analysis.

[0534] System implementation and verification

[0535] The present application has been verified by laboratory and field tests.

[0536] The experimental platform configuration is shown in Table 1:

[0537] Table 1: Experimental platform configuration

[0538]

[0539] The anomaly detection performance comparison results are shown in Table 2:

[0540] Table 2: Performance comparison of different detection methods

[0541]

[0542] The detection performance of different types of anomalies is shown in Table 3:

[0543] Table 3: Detection performance of different types of anomalies (F1 score)

[0544]

[0545] The edge-cloud collaborative computing performance analysis is shown in Table 4:

[0546] Table 4: Edge-cloud collaborative computing performance analysis

[0547]

[0548] The early warning time advance amount comparison is shown in Table 5:

[0549] Table 5: Early warning time advance comparison (minutes)

[0550] Fault type Threshold method Traditional machine learning Deep learning The present method Hot spot development 12.5 28.6 42.3 68.7 Inverter overheating 8.3 19.5 35.2 52.4 Loose connection point 5.6 13.8 25.7 43.6 Array matching anomaly 10.2 22.4 38.6 56.8 Power grid harmonic interference 3.5 9.8 18.3 32.5

[0551] The experimental results show that the abnormal monitoring and early warning method based on edge computing proposed in this patent is superior to the traditional method in various performance indicators. Especially in terms of detection accuracy, response time and early warning ability, the accuracy reaches 96.2%, the response time is only 48ms, the early warning time is advanced by an average of 43-68 minutes, which greatly improves the reliability and safety of the system. The edge-cloud collaborative architecture effectively balances the computing performance and resource consumption, while maintaining good scalability and model updating efficiency.

Claims

1. A method for anomaly monitoring and intelligent early warning in a distributed photovoltaic system, characterized in that, Includes the following steps: S1: High-frequency acquisition of key parameters of the distributed photovoltaic system, including but not limited to the photovoltaic array voltage V, current I, active power P, reactive power Q, temperature T, irradiance Ir, η is the conversion efficiency, environmental parameter vector Env and equipment state parameter θ, forming a state vector S; each parameter in the state vector S is acquired according to its corresponding sampling frequency, and a multi-channel time series data matrix X(t) is output. S={V,I,P,Q,η,T,I r ,Env,θ}; S2: The multi-channel time-series data matrix X(t) undergoes noise filtering and signal normalization, followed by multi-domain feature extraction, including extraction of time-domain, frequency-domain, and time-frequency-domain features, outputting a standardized multi-dimensional feature vector F and time-frequency image features I. gram ; S3: Standardized multidimensional feature vector F and time-frequency image I gram Anomaly detection is performed by autoencoder reconstruction error analysis, convolutional neural network feature classification, and single-class support vector machine. The outputs are then fused and weighted to obtain the anomaly detection result (Decision) and confidence score (Score). anomaly ; S4: Establish a multi-dimensional anomaly assessment model, including four dimensions of assessment: electrical characteristics, environmental conditions, equipment status, and system performance. Input the anomaly detection result Decision, the multi-dimensional feature vector F, and the data of electrical characteristics, environmental conditions, equipment status, and system performance. The assessment results of each dimension are combined to obtain the final assessment result. S5: Based on the final evaluation results obtained in S4 and the multidimensional feature vector F obtained in S2, perform hierarchical classification to identify specific anomaly types, diagnose different types of anomalies, identify typical anomalies, assess the severity of faults, and identify specific anomaly types, including coarse-grained classification C. coarse and fine-grained classification C fine Output the specific anomaly type, i.e., the fine-grained classification result C. fine The severity level provides detailed diagnostic information for early warning decisions.

2. The method for anomaly monitoring and intelligent early warning of a distributed photovoltaic system according to claim 1, characterized in that, It also includes the following steps: S6: Preset warning levels and establish corresponding thresholds, and set corresponding handling measures for different warning levels, targeting the abnormal type C of S5. fine And the severity level (Severity), outputting warning information M warning And response strategies, and based on the early warning transmission strategy, guide operation and maintenance personnel to take corresponding measures.

3. The method for anomaly monitoring and intelligent early warning of a distributed photovoltaic system according to claim 1, characterized in that, It also includes either or both of the following: Construct an edge-cloud collaborative computing scheduling strategy, dynamically allocate computing tasks to the edge or cloud by combining the computing tasks and system resource status of each step, and output an optimized computing resource allocation scheme to ensure efficient system operation; We continuously collect current operating data, evaluate the current model performance, incrementally learn and update model parameters online, and continuously optimize the model to ensure that detection performance improves with the accumulation of data.

4. The method for anomaly monitoring and intelligent early warning of a distributed photovoltaic system according to claim 1, characterized in that, In S1, the state vector is as follows: S={V,I,P,Q,η,T,I r ,Env,θ}(1) In the formula, V and I are voltage and current vectors, respectively; P and Q are active power and reactive power, respectively; η is the conversion efficiency; T is the temperature vector; and I... r Here, θ represents the irradiance, Env represents the environmental parameter vector, and θ represents the device status parameters. Differentiated sampling strategies are designed for different parameter characteristics, and the sampling frequency design is based on the Nyquist theorem: f sample ≥2·f max ·a(4) In the formula, f max The highest frequency component of the signal is typically 5kHz for voltage and current signals and 50Hz for power signals; the safety factor is typically set to 2-5 to prevent spectral aliasing. Each parameter in the state vector S is collected according to its corresponding sampling frequency to form multi-channel time series data, and the multi-channel time series data matrix X(t) is output. X(t)=[V(t1),V(t2),...,V(t n ); I(t1),I(t2),...,I(t n ); P(t1),P(t2),...,P(t n ); Q(t1),Q(t2),...,Q(t n ); n(t1),n(t2),...,n(t n ); T(t1),T(t2),...,T(t n ); And(t1),And(t2),...,And(t n ); Env(t1),Env(t2),...,Env(t n ); θ(t1),θ(t2),...,θ(t n )] In the formula, X(t) is a 9×N matrix, where 9 represents the number of parameter channels and N is the maximum number of sampling points within the time window, determined by the highest sampling frequency. Where T window This refers to the data collection time window.

5. The method for anomaly monitoring and intelligent early warning of a distributed photovoltaic system according to claim 1, characterized in that, In S2, noise filtering: X d =W -1 (T λ (W(X))) (6) In the formula, X is the original multi-channel time-series signal, W represents the wavelet transform operator, and Daubechies wavelet is used; T λ X is the thresholding function, employing soft thresholding; d The signal is the clean signal after noise reduction, and λ is a threshold parameter that is adaptively adjusted according to the noise level. Signal normalization employs Z-score standardization: In the formula, μ is the mean vector of each channel signal, and σ is the standard deviation vector of each channel signal; X norm The standardized signal matrix; Then, multi-domain feature extraction is performed, including extraction of the time domain F. time Frequency domain F freq and time-frequency domain features F time-freq : F={F time ,F freq ,F time-freq } (8) Temporal features include statistical features and morphological features: F time ={mean,std,kurtosis,skewness,peak,rms,crest,...}(9) Where: mean is the signal mean, std is the standard deviation, kurtosis is the kurtosis, skewness is the skewness, peak is the peak value, rms is the root mean square value, and crest is the peak factor; Frequency domain features are obtained through Fast Fourier Transform: F freq ={f c ,bandwidth,harmonic r atio,energy b and,...}(10) In the formula: fc is the center frequency, bandwidth is the frequency band, and harmonic r Oatio is the harmonic ratio. energy b and represents the frequency band energy; Time-frequency domain features are extracted using wavelet packet transform: F time-freq ={energy i ,entropy i ,...,i=1,2,...,n}(11) Where: energy i Entropy is the energy of the i-th frequency subband. i Let i be the entropy value of the i-th frequency sub-band. i = 1, 2, ..., n, where n is the number of sub-bands; Image features are obtained by converting time-series data into a two-dimensional representation: I gram =f transform (X seg )(12) In the formula, X seg For a data segment within a time window, f transform Transformation functions from signals to images, such as time-frequency graphs and recursion graphs; Output standardized multidimensional feature vector F and time-frequency image I gram ; The process of constructing the feature vector F: F=concatenate(F time ,F freq ,F time-freq ) Wherein: F time Calculated by equation (9), it contains 7-dimensional statistical characteristics; F freq The Ftime-freq is calculated by Equation (10) and contains 4-dimensional frequency domain features; the Ftime-freq is calculated by Equation (11) and contains n-dimensional time-frequency features. concatenate represents the feature concatenation operation.

6. The method for anomaly monitoring and intelligent early warning of a distributed photovoltaic system according to claim 1, characterized in that, In S3 The self-encoder reconstruction error analysis is as follows: h=f encoder (X;θ e (13) In the formula, f encoder Let X be the encoder function, and let F be the standardized multidimensional feature vector as input; h is the low-dimensional hidden layer representation vector output by the encoder. The feature vector reconstructed by the decoder; θ e and θ d These are the encoder and decoder parameters, including weights and biases; L reconstruction The reconstruction loss measures the difference between the input and the reconstructed output. Anomaly scoring is based on reconstruction error calculation: In the formula, Score AE (X) represents the autoencoder anomaly score; the larger the value, the more abnormal the anomaly. The convolutional neural network feature classification described above: h conv =CNN(I gram ;θ cnn ) (17) p(y|I gram )=softmax(W·h conv +b) (18) In the formula, I gram The input image has dimensions H×W×C; h conv These are convolutional features, extracted through multiple convolutional layers. θ cnn This is the set of all trainable parameters for a CNN model, including kernel weights and biases. p(y|I gram ) represents the posterior probability distribution of the class of the given input image; W is the classifier weight matrix; b is the classifier bias vector; CNN models employ a lightweight architecture, making them suitable for deployment on edge devices. CNN=Conv(3,16,3,1)-BN-ReLU-MP(2)-Conv(16,32,3,1)-BN-ReLU-MP(2)-FC(128)-FC(n class (19) In the formula, Conv(in,out,k,s) represents a convolutional layer, with parameters being the number of input channels, the number of output channels, the kernel size, and the stride, respectively; BN represents a batch normalization layer; ReLU represents an activation function; MP(k) represents a max pooling layer, with the kernel size as the parameter; and FC(n) represents a fully connected layer, with the number of neurons as the parameter. The single-class support vector machine is used for anomaly detection: In the formula, K(x,x) i ) is the kernel function, α i For Lagrange multipliers, ρ is the bias term, and N is the value of N. s This refers to the number of support vectors. The output results are fused and weighted to obtain the anomaly detection result (Decision) and the confidence score (Score). anomaly : Output process:

1. The three detection models output the following scores: AE (X)(Equation 16), p(y|I) gram (Equation 18), f OCSVM (x)(Equation 20); 2. The outputs of the three models are weighted and fused according to equation (21) to obtain the comprehensive anomaly score. anomaly ; 3. Set the score anomaly The binary decision result Decision is generated according to Equation (22) after comparison with the adaptive threshold.

4. Simultaneously output the confidence score. anomaly The values ​​provide a quantitative basis for subsequent evaluation; Outputs real-time anomaly detection results (Decision and Confidence Score). anomaly This provides a basis for subsequent multidimensional assessments; Score anomaly =w1·Score AE +w2·(1-max(p(y=normal|I gram )))+w3·D OCSVM (21) In the formula, w1, w2, and w3 are weight coefficients that satisfy w1 + w2 + w3 = 1. Performance is verified based on each model. AE Scoring the autoencoder reconstruction error; max(p(y=normal|I gram D represents the maximum predicted probability of the CNN model for the normal category. OCSVM The distance from the sample to the OCSVM decision boundary; Anomaly detection results are obtained by comparing a comprehensive score with an adaptive threshold: In the formula; Decision represents the final anomaly detection decision, with values ​​of {Normal, Anomaly}. Threshold adaptive The threshold is adaptive and dynamically adjusted according to the distribution of historical data. If the overall anomaly score is Score anomaly If the value is less than or equal to the adaptive threshold, it is considered normal; otherwise, it is considered abnormal. The output Decision is the final abnormality detection result. The adaptive threshold is dynamically adjusted based on the distribution of historical data. Threshold adaptive =μ Score +b·s Score (23) In the formula, μ Score and σ Score These are the mean and standard deviation of the scores, respectively, and β is the adjustment coefficient; Outputs real-time anomaly detection results (Decision and Confidence Score). anomaly This provides a basis for subsequent multidimensional assessments.

7. The method for anomaly monitoring and intelligent early warning of a distributed photovoltaic system according to claim 1, characterized in that, S4 specifically refers to, Establish a multidimensional anomaly assessment model, including an electrical characteristic assessment model E. electrical Environmental condition assessment model E environmental Equipment condition assessment model E device System performance evaluation model E performance ; AND comprehensive ={And electrical ,AND environmental ,AND device ,AND performance } (24) Electrical characteristic evaluation model Electrical characteristic assessment focuses on, but is not limited to, anomalies in electrical parameters such as voltage, current, and power. E electrical =f e (V,I,P,Q,PF,THD,...) (25) In the formula, V, I, P, and Q represent voltage, current, active power, and reactive power, respectively, and f e The electrical characteristic evaluation function is based on expert rules and a data-driven approach; PF is the power factor, and THD is the harmonic distortion rate. Environmental condition assessment model Environmental condition assessment considers the impact of environmental factors such as temperature, humidity, and irradiance on photovoltaic systems: E environmental =f env (T,H,I r ,W s ,Dust,...) (26) In the formula, f env For environmental condition evaluation functions, T, H, I r W s Dust represents temperature, humidity, irradiance, wind speed, and dust deposition, respectively. Equipment condition assessment model Equipment condition assessment focuses on the health status of components and inverter equipment: E device =f d (Aging,Hot spot ,Connection,Inverter status ,...) (27) In the formula, Aging represents the degree of component aging, and Hot represents the degree of component aging. spot This indicates the hotspot situation; Connection indicates the connection status; Inverter status Indicates the inverter status; System performance evaluation model System performance evaluation focuses on the overall power generation performance of the photovoltaic system: E performance =f p (η,PR,Yield,...) (28) In the formula, η is the conversion efficiency, PR is the performance ratio, and Yield is the power generation. Fuzzy Logic Comprehensive Evaluation The evaluation results from each dimension are synthesized using a fuzzy logic system to obtain the final evaluation: In the formula, A1, A2, etc., are fuzzy sets of input variables, B i For the fuzzy set of the output variables; The fuzzy reasoning process is based on the Mamdani inference model: In the formula, x represents j For fuzzy set A j membership degree This indicates that the output y is related to the fuzzy set B. i Membership degree; The final evaluation results were obtained through centrifugation: In the formula, y i To output the sampling points, μ(y) i ) represents its membership degree; Output four-dimensional evaluation vector E comprehensive and comprehensive health index E final It provides detailed status information for anomaly type identification; Ecomprehensive=[Eelectrical,Eenvironmental,Edevice,Eperformance] The evaluation results for each dimension are normalized: Final overall health index E final The value range is [0,1], obtained through fuzzy logic reasoning of equation (31), where 0 represents severe abnormality and 1 represents complete normality.

8. The method for anomaly monitoring and intelligent early warning of a distributed photovoltaic system according to claim 1, characterized in that, S5 refers to: Input data: The four-dimensional evaluation vector E output by S4 comprehensive and comprehensive health index E final And the standardized multidimensional feature vector F output by S2 Data Fusion: F enhanced =concatenate(F,E) comprehensive E final ) Hierarchical classification: using F enhanced Perform coarse-grained and fine-grained classification Output: Exception type C fine Severity S51: Hierarchical Classification Recognition Coarse-grained classification uses global features: F global =F enhanced C coarse =f classifier1 (F global ) (32) In the formula, F global The global feature vector containing features across all dimensions is equal to the enhanced feature vector F. enhanced ;F enhanced It integrates the original features F and the evaluation results E. comprehensive Health E final Enhanced features; f classifier1 C is a coarse-grained classifier function used to distinguish between major anomaly classes. coarse The results are coarse-grained classifications, with values ​​of {electrical anomaly, environmental anomaly, equipment anomaly, performance anomaly}. Fine-grained classification uses specific features: F specific =feature selection (F enhanced ,C coarse ) C fine =f classifier2 (F specific |C coarse ) (33) In the formula, F specific Based on the coarse-grained classification result C coarse Select a specific subset of category-specific features; f eature selection The feature selection function is used to select relevant features based on the coarse-grained results; f classifier2 For fine-grained classifier functions, optimize C for specific anomaly types. fine For fine-grained classification results, specific anomaly types include {hot spot failure, component aging, shadow occlusion, etc.}; S52: Diagnosis of different types of abnormalities Design diagnostic models for different types of anomalies: In the formula, X is a dedicated diagnostic function for the i-th type of anomaly; i For the feature subset related to anomaly type i, Θ i For diagnostic model parameters; g i Let i be the feature transformation function for the i-th type of anomaly; S53: Identification of Typical Anomalies Common anomaly types and their characteristic manifestations: Hot spot mode: Pattern hotspot ={T local >T threshold ,ΔT>ΔT threshold ,I string <I normal ,...} (35) In the formula, T local This is a local temperature measurement; T threshold The temperature threshold for hot spot detection is typically 80°C; I string I represents the string current value. norma l represents the string current value under normal operating conditions; Performance degradation: Pattern degradation ={P actual / P rated <the> threshold ,Δη / Δt>slope threshold ,...} (36) In the formula, P actual P represents the actual power output (W); rated Rated power output (W); Δη / Δt is the power attenuation slope; slope threshold This is the threshold for determining performance degradation. Shadow occlusion: Pattern shading ={I string <I normal ,P<P expected when I r >I r,threshold ,I mismatch ,...} (37) In the formula, I string For the affected string current, I normal For normal string current (A), P expected For the desired power based on irradiance, I r For the current irradiance, I r,threshold I is the current difference threshold. mismatch This refers to the mismatch current index. Inverter malfunction: In the formula, V output THDt is the inverter output voltage (V). hreshold η is the voltage harmonic distortion rate threshold, and η is the inverter conversion efficiency. inv,threshold This is the efficiency threshold; S54: Fault Severity Assessment Severity=α1·D feature +α2·T duration +α3·I impact (39) In the formula, D feature T represents the degree to which a feature deviates from its normal value. duration For the duration of the abnormality, I impact The impact of anomalies on system performance is represented by α1, α2, and α3, which are weighting coefficients. Output the specific anomaly type (fine-grained classification result) C fine And the severity level, providing detailed diagnostic information for early warning decisions; Severity assessment integration: Severity=λ1·D feature +λ2·T duration +λ3I impact Where: λ1, λ2, and λ3 are weight coefficients, corresponding to the importance of feature deviation, duration, and degree of influence, respectively; D feature To indicate the degree to which a feature deviates from its normal value, based on F enhanced Calculation; T duration The duration of the anomaly is statistically derived from the time-series data X(t); I impact The impact of the anomaly on system performance is quantified by combining the Efinal health index; Severity is the score for the severity of the fault, with a value range of [0,1]. Output in a uniform format: The final output is structured data; Output S5 ={ anomaly type :C fine , severity score :Severity, confidence:max(p(C fine |F enhanced )), feature importance :ranking(F enhanced ), temporal info :T duration } Ensures the output C of S5 fine Severity and the output of the previous steps are kept consistent in data flow and logic, forming a complete chain of anomaly diagnosis.

9. The method for anomaly monitoring and intelligent early warning of a distributed photovoltaic system according to claim 1, characterized in that, S6 refers to: S61: Preset warning levels and set corresponding thresholds. The tiered early warning mechanism classifies the system status into five levels based on the anomaly assessment results: normal, attention, warning, alarm, and danger. In the formula, L1, L2, L3, and L4 are the warning thresholds; E final This is a comprehensive health index, with values ​​ranging from [0,1]. Level represents the warning level, with values ​​{Normal, Attention, Warning, Alarm, Danger}. S62: Establish corresponding handling measures for different levels of early warning. Define the early warning strategy matrix Strategy(Level,Type)=Action specific (41) In the formula: Strategy is the early warning strategy matrix, which defines the handling measures corresponding to different levels and types. Level is the early warning level, Type is the exception type, and Action is the exception type. specific For targeted response actions; S63: Early Warning Information Generation Early warning information generation model: M warning ={Level,Type,Location,Time,Severity,Suggestion,Confidence} (42) In the formula, Level is the warning level, Type is the anomaly type, Location is the anomaly location, Time is the detection time, Severity is the severity, Suggestion is the handling suggestion, and Confidence is the confidence level. S64: Early Warning Transmission Strategy Early warning transmission priority is based on importance assessment: P transmission =f t (Level,Type,Receiver) (43) In the formula, f t For the early warning propagation strategy function, P transmission To prioritize the transmission of early warnings, the Receiver is the recipient; Output complete early warning information M warning And response strategies to guide operations and maintenance personnel to take appropriate measures.

10. The method for anomaly monitoring and intelligent early warning of a distributed photovoltaic system according to claim 3, characterized in that, Construct an edge-cloud collaborative computing scheduling strategy, dynamically allocate computing tasks to the edge or cloud based on the computing tasks and system resource status at each step, and output an optimized computing resource allocation scheme to ensure efficient system operation, including: S71: Task Scheduling Strategy The edge-cloud collaborative computing architecture adopts a layered design, dynamically allocating tasks based on computing needs and resource conditions: Task Allocation =f scheduler (Task Property ,Resource State ) (44) In the formula, Task Allocation Assign results to tasks and specify the execution location for each task; f scheduler This is a task scheduler function based on a heuristic algorithm; Task Property Including computational complexity, time sensitivity, and resource State This includes computing power, load status, and resource information; S72: Multi-objective optimization Task scheduling strategies are based on multi-objective optimization: minJ=w1·T processing +w2·E consumption +w3·B usage (45) s.t.T processing ≤T deadline (46) AND consumption ≤E budget (47) Resource Usage ≤Resource Capacity (48) In the formula, T processing For processing time, E consumption For energy consumption, B usage For bandwidth usage, w1, w2, and w3 are weighting coefficients; T deadline E is the deadline for the task. budge t represents the energy consumption budget (J), and Resource Capacity This represents the maximum resource capacity. S73: Model Compression and Deployment Lightweight model deployment on the edge side employs model compression technology: Model edge =f compress (Model original ,Compression Ratio ) (49) Compression methods include pruning, quantization, and knowledge distillation: Model pruned =f prune (Model original ,Pruning Strategy ) (50) Model quantized =f quantize (Model pruned ,Bit Width ) (51) Model distilled =f distill (Model teacher ,Model student ) (52) In the formula, Model edge For the edge-side lightweight model, f compress Model is a model compression function. original For the original complete model, Compression Ratio For compression ratio, Modelp runed For the pruned model, f prune Pruning is a pruning function. Strategy For pruning strategies, Model quantized For the quantized model, f quantize For the quantization function, Bit Width To quantize the bit width, Model distilled For the distillation model, f distill For the knowledge distillation function, Model teacher For the teacher model, Model student For student models; S74: Edge Capability Assessment Edge device computing power assessment model: C edge ={CPU:f cpu ,RAM:m ram ,Storage:s storage ,Power:p max } (53) In the formula, C edge To score the overall computing power of edge devices, f cpu m is a CPU performance metric. ram For memory capacity, s storage For storage capacity (GB), p max This represents the maximum power consumption. S75: Data Interaction Strategy Edge-cloud data interaction strategy: D upload =f filter (D raw ,Policy upload ) (54) Policy upload ={Rule1,Rule2,...,Rule n } (55)Rule i :IF Condition i THEN Action i (56) In the formula, D upload For filtered data that needs to be uploaded to the cloud, f filter D is a data filtering function. raw For the raw collected data, Policy upload For a set of data upload strategies, the Rule i The i-th upload rule, Condition i As the triggering condition for the i-th rule, Action i This refers to the action to be executed for the i-th rule. Output optimized computing resource allocation schemes to ensure efficient system operation; Continuously collect current operational data, evaluate the current model performance, incrementally learn and update model parameters online, and continuously optimize the model to ensure that detection performance improves with data accumulation. Specifically, this includes: S81: Incremental Learning Update The online model update mechanism is based on incremental learning technology to achieve continuous model optimization. I t+1 =Θ t +DTH t (57) DTH t =f update (D new ,I t ,Performance t ) (58) In the formula, Θ t Let D be the model parameters at time t. new For the newly collected data, f update Update function for parameters, Performance t For current performance evaluation; S82: Incremental Learning Loss Incremental learning strategy design: L incremental =L new +λ·L memory (59) In the formula, L incremental For incremental learning total loss function, L new For the loss on new data, L memory To prevent catastrophic forgetting and memory retention loss, λ is a parameter for balancing old and new knowledge; S83: Knowledge Distillation Update Knowledge distillation is used for model updates: In the formula, L CE To calculate the true label y and the predicted result Cross-entropy loss between, L KL Let z be the KL divergence. t and z t-1 These are the outputs of the current model and the historical model, respectively, where τ is the temperature parameter and α is the balance coefficient; S84: Adaptive threshold adjustment Adaptive threshold adjustment strategy: Threshold t+1 =Threshold t +γ·(FAR t -Target FAR ) (61) In the formula, FAR t Target is the current false positive rate. FAR γ is the target false alarm rate, and γ is the adjustment coefficient. S85: Performance Evaluation Metrics Feature importance assessment and selection: Importance(F i )=Information Gain (Class,F i ) (62) IG(Class,F i )=H(Class)-H(Class|F i ) (63) In the formula, Importance(F) i ) is the importance score for the i-th feature, Information Gain Information gain measures the classification ability of a feature; F i Let IG(Class,F) be the i-th feature, and Class be the class label. i H(Class) is the information gain of feature Fi on class Class; H(Class|F) is the class entropy. i ) is a known feature F i Category entropy under certain conditions; Model performance evaluation metrics: In the formula, TP, FP, and FN represent the number of true positive, false positive, and false negative samples, respectively; Precision is the accuracy rate, the proportion of samples correctly predicted as abnormal; Recall is the recall rate, the proportion of actual abnormalities correctly identified; F 1Score is the harmonic mean of precision and recall; AUROC is the area under the ROC curve, which measures the overall performance of the classifier; t is the classification threshold parameter; TPR is the true positive rate, and FPR is the false positive rate.

Images