State early warning method for photovoltaic power station inverter of integrated energy dispatching system

Abstract

The invention provides a state early warning method for a photovoltaic power station inverter of a comprehensive energy dispatching system, and the method comprises the steps: reading the historical time series data of the current change of the inverter in a specific time period and the weather data of the same period, and obtaining a plurality of intrinsic mode functions with the original specific characteristics of the current of the inverter; and taking the intrinsic mode function and the weather data as training data, predicting the current of the inverter in a specific time period based on the trained recurrent neural network model, searching and comparing a predicted value with a preset inverter current early warning threshold value, and judging whether the state of the inverter needs to be early warned in the time period or not. According to the method, the scheme of combining the variational mode decomposition with the recurrent neural network is adopted, the intrinsic mode function in the current data of the inverter is extracted through the VMD, the weather factors are combined, the inverter is input into the RNN for deep learning, and the prediction precision of the future state of the inverter is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of integrated energy dispatching and photovoltaic power generation, and in particular to an inverter state early warning method for a photovoltaic power station in an integrated energy dispatching system. Background Art

[0002] In the field of integrated energy system scheduling, effective management of distributed renewable energy resources is crucial, especially in photovoltaic power stations, where inverter health management and early warning are extremely critical to ensuring stable system operation and timely maintenance.

[0003] The inverter is responsible for converting direct current into alternating current, and its performance directly affects the efficiency of the photovoltaic power station and even the entire energy system. However, the existing early warning and management methods have the following main problems:

[0004] 1. Over-reliance on manual labor: Traditional methods rely on a large amount of human resources for regular inspection and maintenance, which cannot achieve early prevention and can only respond passively after problems appear;

[0005] 2. Lack of predictive capabilities: Most existing automation systems only monitor the current status and lack predictions of future operating conditions, making it impossible to detect potential problems in advance;

[0006] 3. Insufficient forecast accuracy: Existing forecasting systems are mostly based on simplified statistical or mathematical models, which fail to fully consider the weather conditions and equipment complexity in actual operation, resulting in inaccurate forecast results.

[0007] Based on this, there is an urgent need for an efficient inverter status warning method that can accurately predict the inverter status to reduce manual dependence and provide early warning. Summary of the invention

[0008] In view of the shortcomings of the prior art, the present invention aims to provide a photovoltaic power station inverter status early warning method for an integrated energy dispatching system, which adopts the technology of recurrent neural network (RNN) and superimposed variational mode decomposition (MVMD) to solve the problems raised in the above background technology.

[0009] In order to achieve the above object, the present invention is implemented through the following technical solutions: a photovoltaic power station inverter state early warning method of an integrated energy dispatching system, comprising:

[0010] S1. Read the historical time series data and weather data of the same period that characterize the inverter current change within a specific period, clean the historical time series data and weather data, fill in missing values ​​to ensure the quality of the data, and use the variational mode decomposition method to process the series data to obtain several intrinsic mode functions IMFs with original specific characteristics of the inverter current;

[0011] S2, taking the intrinsic mode function IMF and weather data as training data, and inputting them into the recurrent neural network model for training;

[0012] S3. Predict the inverter current in a specific period of time based on the trained recurrent neural network model, and compare the predicted value with the preset inverter current warning threshold to determine whether it is necessary to issue a warning for the inverter status in this period of time.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] 1. In order to solve the problem of response lag caused by reliance on manual inspection and maintenance in the prior art and the limitation that the automation system can only monitor in real time but cannot predict the future state, the present invention adopts a solution of variational mode decomposition (VMD) combined with recurrent neural network (RNN). First, the historical time series data and weather data are cleaned and missing values ​​are filled to ensure the quality of the data. Secondly, feature engineering is performed on the intrinsic mode function IMF and weather data to extract the most informative features in the intrinsic mode function IMF, reduce the dimension and improve the model training efficiency. Finally, the intrinsic mode function in the inverter current data is extracted by VMD, and combined with weather factors, it is input into RNN for in-depth learning to improve the prediction accuracy of the future state of the inverter.

[0015] 2. In order to solve the problem of insufficient accuracy caused by simplified models in the existing field of integrated energy system scheduling, the present invention uses a deep learning model to learn and master the operating rules of complex and nonlinear equipment. At the same time, by training a recurrent neural network model, it effectively predicts the working state changes of the inverter, realizes early warning, reduces inverter alarms and damage, and enhances the operating stability of the photovoltaic power station;

[0016] 3. In addition to the historical operating current data of the inverter, the model learning data used in the present invention fully considers the impact of weather factors including sunshine intensity and temperature on the operation of the inverter in the photovoltaic power station. The historical data of these two parameters are input into the recurrent neural network model together with the modal function decomposed by the above-mentioned VMD method as model training data, which solves the problem of inaccurate prediction results caused by sparse training sample data. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are only for illustrative purposes and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0018] Figure 1 It is a schematic diagram of the overall architecture of a photovoltaic power station inverter working state early warning method proposed in an embodiment of the present invention;

[0019] Figure 2A This is a schematic diagram of time series data representing inverter current changes within 5 days proposed in one embodiment of the present invention;

[0020] Figure 2B It is a summary schematic diagram of 5-day inverter current time series data intercepted according to an embodiment of the present invention;

[0021] Figure 2C It is a schematic diagram of temperature changes corresponding to the summary of inverter current time series data within a 5-day period proposed in an embodiment of the present invention;

[0022] Figure 2D It is a schematic diagram of the change of the illumination curve corresponding to the inverter current time series data collected within a 5-day period proposed in one embodiment of the present invention;

[0023] Figure 3 It is a schematic diagram of each mode signal and residual obtained when the variational mode decomposition method is used to process the sequence data in one embodiment of the present invention;

[0024] Figure 4 A schematic diagram of a recurrent neural network combined with a variational mode decomposition method proposed in one embodiment of the present invention;

[0025] Figure 5 A schematic diagram of the prediction effect using a recurrent neural network model proposed in one embodiment of the present invention. DETAILED DESCRIPTION

[0026] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific implementation modes and drawings are only exemplary descriptions of the technical solution of the present invention, and should not be regarded as the whole of the present invention or as a limitation or restriction to the technical solution of the present invention.

[0027] The present invention is further described in detail below with reference to the accompanying drawings, but is not intended to limit the present invention.

[0028] As an understanding of the technical concept and implementation principle of the present invention, the present invention proposes a photovoltaic power station inverter state early warning method for a comprehensive energy dispatching system. The main purpose is to accurately predict the operating state of the inverter in the photovoltaic power station. It should be noted that this state is affected by two key factors:

[0029] 1. Interaction between inverter and power station equipment: It is understandable that in the process of interaction between the inverter and the nonlinear impedance in the power station and other inverters, due to the presence of many nonlinear devices in the power station, it is a huge challenge to build an accurate and stable inverter model. The model construction is complex and full of uncertainty.

[0030] 2. Dynamic impact of weather conditions: It is understandable that weather conditions such as sunshine intensity and temperature have a direct impact on the power generation of photovoltaic power stations. These weather factors make it difficult to accurately predict the operating power of the inverter through traditional mathematical and physical models. In existing numerical analysis, it is difficult to integrate weather factors as an exact variable, so it is impossible to accurately evaluate its specific impact on the operation of the power station.

[0031] Therefore, the present invention adopts Recurrent Neural Networks (RNN) and superimposed Variational Mode Decomposition (MVMD) technology to study the current time series data of the inverter for the overcurrent problem, and further realizes the accurate prediction of its future state, so as to judge the future state changes of the inverter, make early warning preparations in advance, and realize high-precision operation and management of the integrated energy dispatching system and photovoltaic power station.

[0032] like Figure 1 , Figure 4 As shown, as an embodiment of the present invention, a method for early warning of inverter status of a photovoltaic power station in a comprehensive energy dispatching system is proposed, comprising the steps of:

[0033] S1. Read the historical time series data and weather data of the same period that characterize the inverter current change within a specific period, clean the historical time series data and weather data, fill in the missing values ​​to ensure the quality of the data, and use the variational mode decomposition method to process the sequence data to obtain a number of intrinsic mode functions IMFs with original specific characteristics of the inverter current. In specific implementation, the weather data of the same period at least includes the temperature history data and sunshine intensity history data that affect the operation of the inverter in the photovoltaic power station;

[0034] S2. Perform feature engineering on the intrinsic mode function IMF and weather data to extract the most informative features in the intrinsic mode function IMF, reduce the dimension and improve the efficiency of subsequent model training. Then, the intrinsic mode function and weather data are used as training data and input into the recurrent neural network model for training. It performs the following operations:

[0035] Normalizing the continuous moment input values ​​obtained from the intrinsic mode function and weather data, and constructing a corresponding time series data set, which is input into a recurrent neural network model;

[0036] Perform forward propagation calculations to calculate the hidden state and output at the current moment based on the recurrent neural network model using the input values ​​at the current moment and the previous moment;

[0037] Define a loss function to measure prediction accuracy, and perform a backpropagation algorithm to calculate the gradient of the loss function relative to the parameters of the recurrent neural network model. Use an optimization algorithm such as gradient descent to update the model parameters based on this gradient.

[0038] Through iterative training, the forward propagation, loss calculation, back propagation and parameter update steps are repeated until the performance of the recurrent neural network model is stable on the validation set or the set number of iterations is reached.

[0039] S3. Predict the inverter current in a specific period of time based on the trained recurrent neural network model, and compare the predicted value with the preset inverter current warning threshold to determine whether it is necessary to issue a warning for the inverter status in this period of time.

[0040] In one embodiment of the present invention, the variational mode decomposition (VMD) method is used as a non-recursive signal decomposition method to process the historical time series data of the inverter current. The obtained data is a series of modal signals with specific frequency characteristics. These modal signals can reflect the different frequency components of the original current data. At the same time, a residual value is obtained during the processing, which reflects the difference between the sum of all modal signals and the original signal, such as Figure 1 As shown, in this embodiment, taking four mode signals as an example, the data in this time period is processed by the variational mode decomposition algorithm. This method can help extract useful information from the inverter current data and provide a basis for subsequent recurrent neural network modeling and prediction.

[0041] Based on the above technical concept, the variational mode decomposition method is used to process the sequence data to obtain several intrinsic mode functions IMF with original specific characteristics of the inverter current. The specific implementation includes:

[0042] S1-1. Each intrinsic mode function IMF is defined as an amplitude modulation frequency modulation function, expressed as:

[0043] u k (t) = A k (t)cos(φ k (t)) (1)

[0044] The purpose is to capture the intrinsic pattern of the current signal, where u k (t) is represented as the kth IMF obtained by decomposition. At the same time, it is necessary to apply Hilbert transform to each decomposed IMF to decompose the sequence data in the intrinsic mode of each current signal according to the following formula, so as to further analyze the frequency characteristics of each mode:

[0045]

[0046] Wherein, πt represents the phase term related to time; j represents the imaginary unit; δ(t) represents the function of time t, which is related to the phase of the signal.

[0047] S1-2. For each mode, the spectrum of the mode is modulated to the baseband using the exponential tuning operator through the following equation (3) to simplify the spectral characteristics of the modal signal and make it easier to process and analyze:

[0048]

[0049] S1-3, the signal is demodulated by Gaussian smoothing technology to achieve bandwidth estimation, and based on the historical time series data I(t) of inverter current changes, combined with the mode set {u k} and the center frequency set ω k , construct a constrained variational problem. In the specific implementation, the Gaussian smoothing technique is used to demodulate the signal by calculating the square root of the L2 norm gradient of the signal. This is a smoothing process of the signal. It can be understood that this smoothing helps to estimate the bandwidth of each IMF, that is, the distribution width of the signal frequency component, so as to achieve the purpose of understanding the signal characteristics and performing effective decomposition.

[0050]

[0051] Where I(t) represents the historical time series data of inverter current change; the mode set {u k}={u1,u2…u N};ω k Represented as u k The kth center frequency, ω k ={ω1,ω2…ω N}; * represents the convolution operation; δ(t) represents the impulse function.

[0052] Based on the above technical concept, it should be noted that since the original constrained variational problem constructed above contains some constraints, such as the non-negativity of signal frequency or constraints of a specific form, these constraints are difficult to handle in the subsequent optimization process. Therefore, the present invention further proposes the steps of:

[0053] S1-4. By introducing the Lagrangian function and the quadratic penalty term, the aforementioned constraints are converted into penalty terms and added to the objective function, thereby converting the problem into an unconstrained problem and simplifying the optimization process. The process is as follows:

[0054] First, define the Lagrangian function of the original constrained variational problem, which combines the objective function, constraints and the corresponding Lagrangian multipliers;

[0055]

[0056] In the formula, L(u k ,ω k ,λ) is represented as Lagrangian function, λ is represented as Lagrangian multiplier; α is represented as penalty term coefficient; It is expressed as a complex exponential term, which is used to represent the phase or frequency characteristics of a signal, where j is an imaginary unit;

[0057] Then, quadratic penalty terms are added to the Lagrangian function, which impose penalties on the objective function when the constraints are violated; then, the coefficients of the penalty terms are adjusted to ensure that the solution satisfies the original constraints; finally, numerical optimization methods are used to optimize the unconstrained objective function in the modified Lagrangian function, usually using numerical optimization methods such as gradient descent, to find the optimal solution to the problem.

[0058] In one embodiment of the present invention, when the unconstrained objective function in the modified Lagrangian function is optimized using a numerical optimization method, the following steps need to be performed:

[0059] S1-5, by using the modal function u k and the center frequency ω k The functional is updated in both directions to find the parameter value that minimizes the unconstrained objective function, which is directly related to the objective of the original constrained variational problem. The process is as follows:

[0060] First, for all ω ≥ 0, we have:

[0061]

[0062]

[0063] Where n is the number of iterations; (^) is the Fourier transform;

[0064] Second, when the convergence criterion is satisfied When , the iteration is completed.

[0065] It can be understood that, according to the above decomposition method, each mode signal after the inverter current is decomposed can be obtained, and the difference norm between the original function (original modal signal) and the decomposition function (all intrinsic mode functions IMF) can be obtained, that is, the residual value, which represents the effect of comparing all the modal functions together with the original signal, such as Figure 3 As shown in FIG. 1 , the signal diagrams and residual diagrams of each mode obtained after variational mode decomposition are shown. In this embodiment, for the convenience of explanation, the inverter current time series diagram of the photovoltaic power station for five consecutive days from June 10 to June 14 is selected as an example, which is input into the variational mode decomposition algorithm. The number of mode functions obtained by decomposition is set to 4, and the images of each mode function and residual value function are shown in FIG. Figure 3 As shown in the figure, in the variational mode decomposition method, the decomposed mode function extracts the implicit characteristics of the original inverter current very well, and each mode function can be regarded as a harmonic signal obtained after the spectrum analysis of the original complex non-periodic signal. Therefore, by calculating the residual value of the original mode signal and the sum of all intrinsic mode functions IMF, the decomposition quality can be evaluated. In specific implementation, if the residual value is smaller, it means that the intrinsic mode function IMF maximizes the approximation of the original mode signal; otherwise, it is necessary to readjust the parameters of the variational mode decomposition method and decompose again.

[0066] As an embodiment of the present invention, since in a traditional neural network model, from the input layer to the hidden layer and then to the output layer, the layers are fully connected but the nodes between each layer are disconnected, which leads to the limited processing ability of ordinary neural networks for many prediction problems. Therefore, the recurrent neural network adopted in the present invention shares weights in different parts of the network model so that the model can be extended to samples of different styles, generalize very long sequences and infer them through memory ability. After training, the current value of the inverter in the future period can be directly predicted to determine whether it needs early warning.

[0067] When training the recurrent neural network model, the time series data set containing the above modal functions and weather data needs to be divided into a training set and a test set, such as Figure 2A As shown, it is the inverter historical working current time series diagram and the weather historical data during the period, including the light intensity time series diagram (such as Figure 2D ) and temperature time series diagram (such as Figure 2C ), in this embodiment, for the convenience of explanation, only the time series diagrams of each parameter for five days from June 10, 2024 to June 14, 2024 are selected (such as Figure 2B), and in actual data-driven model training, data of several months or even years is needed as training data to improve the prediction accuracy of the prediction model. After selecting the data, the training set is set to account for 70% and the test set data is set to account for 30%, and the training set is used to train the recurrent neural network model. The trained model is used to predict the test set data, and the mean square error, root mean square error, and mean absolute error indicators are used to evaluate the prediction accuracy of the recurrent neural network model. The calculation formula is as follows:

[0068]

[0069]

[0070]

[0071] In the formula, N represents the number of samples; yi represents the actual value; Expressed as predicted value.

[0072] The training set prediction results, training set errors and test set prediction results, test set errors are all as in this embodiment. Figure 5 As shown in the figure, the training set error and the test set error were calculated respectively. The mean square error, root mean square error and mean absolute error of the training set were 18.7393, 4.3289 and 5.1104 respectively; the mean square error, root mean square error and mean absolute error of the test set were 88.2996, 9.3968 and 4.1434 respectively. The reason for the large mean square error is that the predicted values ​​at several times have large deviations, which affects the overall accuracy. From the perspective of absolute error, this model still has a relatively considerable accuracy in predicting current. When the model training effect meets the standard, the model can be used to directly predict the current value of the inverter in the future period to determine whether it needs early warning.

[0073] In one embodiment of the present invention, the specific implementation methods of cleaning historical time series data and weather data and filling missing values ​​to ensure the quality of data include:

[0074] First, define an ordered data set I and W to represent the historical time series data set of inverter current changes and the weather data set of the same period, respectively. For the ordered data set I, each element is a pair of (t i ,I(t i )), t i Represented as a timestamp; I(t i ) is represented by the inverter current value at the corresponding time; for the ordered data set W, each element of it is (t i ,T(t i ),S(t i )), T(ti ) is represented by time t j Temperature; S(t i ) is the sunshine intensity;

[0075] Secondly, using a data frame D frame , each row is assumed to represent a time point t i Each column of the data record represents a specific measurement parameter:

[0076]

[0077] Again, in the data frame D frame Data cleaning is performed in the process, such as using the standard deviation method to identify and process to remove outliers and using the interpolation method to fill in the data, so as to ensure that the historical time series data of inverter current changes and weather data obtained are accurate at each time point t i All data are recorded synchronously to ensure the quality of the data.

[0078] In specific implementation, the predicted value of the missing point is calculated by weighted averaging the values ​​of the known data points, and the inverter current I(t i ) and weather data W(t i ) is estimated, and the weight of the data point is determined by the Gaussian function to reflect the decreasing effect of distance. Over time t i The time difference d between the point t to be inserted i The principle of exponential increase and exponential decay distinguishes the present invention from traditional neural network models. It effectively utilizes and strengthens the spatial correlation of time series data, and ensures the continuity and integrity of the sample data input into the training model in the time series by making reasonable estimates of missing data, thereby ensuring the quality of the data and outputting accurate prediction results.

[0079] At time point t i The calculation formula for data interpolation is:

[0080]

[0081] In the formula, S t It is represented as the historical time series data and weather data of the same period read within a specific period, at time point t i The comprehensive observation vector at i ) represents the historical time series data and weather data of the same period read within a specific period, at time point t i The known observation vector at ; n is represented by the data frame D frame When performing data cleaning, the actual number of data points that need to be interpolated and supplemented; It is represented by a comprehensive weight function that is set and adjusted according to the correlation between the inverter current historical time series data and the weather data, where ∈1 and ∈2 represent the weight coefficients of the inverter current historical time series data and the weather data, respectively, which are adjusted according to the actual changes; σ I and σ W are control parameters represented as weight coefficients of inverter current historical time series data and weather data respectively.

[0082] Based on the above technical concept, considering that weather data makes it difficult to accurately predict the operating power of the inverter through traditional mathematical and physical models, therefore, this embodiment proposes to use interpolation method to fill in the data.

[0083] In order to further ensure that the filled data is continuous and complete, based on the above-mentioned interpolation method to fill the data concept, the present invention further proposes to use the MK test to identify the key time points in the data filling process, so as to select the data involved in the interpolation and fill the inverter current and weather data in the missing time period, thereby ensuring the continuity and integrity of the data, thereby improving the prediction accuracy and providing support for subsequent further analysis and model training.

[0084] In one embodiment of the present invention, the specific implementation of applying the MK test to determine the data involved in interpolation includes:

[0085] First, prepare the previously defined time series data, the historical time series data set I of inverter current changes, the weather data set W of the same period, and the data framework D frame , waiting for MK inspection;

[0086] Secondly, the MK check is performed on the historical time series data set I of the inverter current change, in order to detect the trend in the time series. It should be noted that the process of performing MK check on the data set I usually involves constructing an order series K. j , the order is K j It records how many data points are before the time point than I j In specific implementation, for each recorded data point in the time series data set I (except the first one), check its previous current value. If the current current value is greater than all previous values, a positive indicator q is assigned. i , otherwise let q i = 0, add up all the indices to get the order column K j ;

[0087] Again, for each order column K j , calculate each order column K according to half of the number of data points in the time series data set I that are greater than the current value j The expected value E(Kj ), calculate each order column K according to a fixed formula based on the length of the time series data set I and the current position j The variance V(K j );

[0088] From this, use the order column K j , expected value E(K j ) and variance V(K j ) Calculate the statistic P in the MK test fj , to detect the development trend of the time series data set I; at the same time, the time series data set I is reversed, and the above steps are repeated to obtain the reverse statistics P of the time series data set I FJ ;

[0089] Finally, by drawing P FJ and P fj , find the point where the sudden change in the trend in the time series intersects, and use it as the point where the trend of the time series data set I suddenly changes. After identifying the trend sudden change point, start from the current point, look for data for the previous U days and the previous V days, and complete the missing time period. The historical time series data of the inverter current change and the weather data of the same period are filled.

[0090] Example: Select the parameters for the five days from June 10, 2024 to June 14, 2024, and perform trend analysis and data interpolation settings using the inverter current, temperature, and sunshine intensity data based on the MK (Mann-Kendall) test steps. Suppose the data frame D formed by the parameters for the five days is frame as follows:

[0091]

[0092] For time t1 (June 10, 2024), the current I(t1) is 150A, the temperature T(t1) is 34°C, and the sunshine intensity S(t1) is 4000lx. The parameters for time t2, t3, t4, and t5 are as shown in the above data frame D. frame As shown;

[0093] Using the inverter current data I to use the MK test, first construct the order column K j , and then calculate the statistic P in the MK test fj . According to the order of Then, since I(t1)>I(t2), K2=0; since I(t1)>I(t3) and I(t2)<I(t3), K3=1; since I(t1)>I(t4) and I(t2)<I(t4) and I(t3)<I(t4), K4=1; since I(t1)>I(t5) and I(t2)<I(t5) and I(t3)>I(t5) and I(t4)>I(t5), K5=3, and so on.

[0094] Use the calculated K j Value, according to the formula And statistical software to calculate the statistic P fj As an indicator of the statistical significance of the trend, it should be noted that if P fj is in reverse order, then P FJ The calculation is based on K j At this point, assuming that a possible trend mutation point is found, P is calculated by the above formula fj and P FJ Identify trend mutation points. It is understandable that in this example, in order to simplify the visual inspection process, it is set to directly select June 12, 2024 (time t3) as the missing data, then based on this mutation point (in practice P fj The calculation should take into account its positive and negative values, a positive number indicates an upward trend, and a negative number indicates a downward trend). The data from June 10 (time t1) to June 11 (time t2) are selected as Frontier (data from the previous U days), and the data from June 13 (time t4) to June 14 (time t5) are used as Rear (data from the next V days) to interpolate the current value at time t3. If no mutation point is found within U or V days, the data within X days before and after the missing data are selected.

[0095] Using the selected data, the inverter current on June 12 is interpolated by the weighted average method, in which the weight is determined by the time point t i ; The time difference between the missing time point and the missing time point is determined, and the weight function usually takes the form of a Gaussian function to ensure that the data points closer to the missing point have a higher weight in the interpolation ω i Bigger: Calculate ω1, ω2, ω4 and ω5 in this way, and the calculated The interpolated current is:

[0096] Based on this, a reasonable estimate can be obtained to fill in the missing data of the inverter current on June 12, while also ensuring the continuity and integrity of the data.

[0097] In one embodiment of the present invention, the specific implementation method of feature engineering of intrinsic mode function IMF and weather data includes at least: basic feature extraction: in the specific implementation, key original data features are identified, including the amplitude and frequency of inverter current changes and the basic statistical characteristics of weather parameters; time series feature extraction: in the specific implementation, the amplitude and frequency of inverter current changes and the dynamic behavior in weather parameter data are analyzed to identify their trends, periodicity and seasonal patterns; advanced feature construction feature extraction: in the specific implementation, by combining basic features and time series features, new feature combinations that can reveal complex patterns of data are created, thereby providing a more accurate and efficient training basis for the model.

[0098] The technical scope of the present invention is not limited to the contents in the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of ​​the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. A photovoltaic power station inverter status early warning method for an integrated energy dispatching system, characterized in that: Includes steps: S1. Read the historical time series data and weather data of the same period that characterize the inverter current change within a specific period, clean the historical time series data and weather data, fill in missing values ​​to ensure the quality of the data, and use the variational mode decomposition method to process the series data to obtain several intrinsic mode functions IMFs with original specific characteristics of the inverter current; S2, taking the intrinsic mode function IMF and weather data as training data, and inputting them into the recurrent neural network model for training; S3. Predict the inverter current in a specific period of time based on the trained recurrent neural network model, and compare the predicted value with the preset inverter current warning threshold to determine whether it is necessary to issue a warning for the inverter status in this period of time.

2. The inverter status early warning method according to claim 1, characterized in that: The weather data for the same period at least includes the historical temperature data and the historical sunshine intensity data that affect the operation of the inverter in the photovoltaic power station. When the historical time series data and weather data are cleaned and the missing values ​​are filled, the following operations are performed: First, define an ordered data set I and W to represent the historical time series data set of inverter current changes and the weather data set of the same period, respectively. For the ordered data set I, each element is a pair of (t i ,I(t i )), t i Represented as a timestamp; I(t i ) is represented by the inverter current value at the corresponding time; for the ordered data set W, each element of it is (t i ,T(t i ),S(t i )), T(t i ) is represented by time t i Temperature; S(t i ) is the sunshine intensity; Secondly, using a data frame D frame , assuming that each row represents a time point t i Each column of the data record represents a specific measurement parameter; Finally, in the data frame D frame Data cleaning is performed in the process of filling data using interpolation method. The process is as follows: The predicted value of the missing point is calculated by weighted averaging the values ​​of the known data points, and the inverter current I(t i ) and weather data W(t i ) is estimated according to the weight function ∈ (di) Over time t i The time difference d between the point t to be inserted i The principle of exponential decay is followed, and the weights of data points are determined by a Gaussian function to reflect the decreasing effect of distance.

3. The inverter status early warning method according to claim 2, characterized in that: At time point t i The calculation formula for data interpolation is: In the formula, S t It is represented as the historical time series data and weather data of the same period read within a specific period, at time point t i The comprehensive observation vector at i ) represents the historical time series data and weather data of the same period read within a specific period, at time point t i The known observation vector at ; n is represented by the data frame D frame When performing data cleaning, the actual number of data points that need to be interpolated and supplemented; It is represented by a comprehensive weight function that is set and adjusted according to the correlation between the inverter current historical time series data and the weather data, where ∈1 and ∈2 represent the weight coefficients of the inverter current historical time series data and the weather data, respectively, which are adjusted according to the actual changes; σ I and σ W are control parameters represented as weight coefficients of inverter current historical time series data and weather data respectively.

4. The inverter status early warning method according to claim 2, characterized in that: When the interpolation method is used for data filling, the MK test is used to identify the key time points in the data filling process, so as to select the data involved in the interpolation, thereby ensuring the continuity and integrity of the data, and performing the following operations: First, prepare the historical time series data set I of inverter current changes, the weather data set W of the same period, and the data framework D frame , waiting for MK inspection; Secondly, the historical time series data set I of the inverter current change is subjected to MK check. For each data point except the first one in the time series data set / , its previous current value is checked. If the current current value is greater than all previous values, a positive indicator q is assigned. i , otherwise let q i = 0, add up all the indices to get the order column K j ; Again, for each order column K j , calculate each order column K according to half of the number of data points in the time series data set I that are greater than the current value j The expected value E(K j ), calculate each order column K according to a fixed formula based on the length of the time series data set I and the current position j The variance V(K j ); From this, use the order column K j , expected value E(K j ) and variance V(K j ) Calculate the statistic P in the MK test fj , to detect the development trend of the time series data set I; at the same time, the time series data set I is reversed, and the above steps are repeated to obtain the reverse statistics P of the time series data set I FJ ; Finally, by drawing P FJ and P fj , find the point where the sudden change in the trend in the time series intersects, and use it as the point where the trend of the time series data set I suddenly changes. After identifying the trend sudden change point, start from the current point, look for data for the previous U days and the previous V days, and complete the missing time period. The historical time series data of the inverter current change and the weather data of the same period are filled.

5. The inverter status early warning method according to claim 1, characterized in that: In S1, when the sequence data is processed using the variational modal decomposition method, the obtained data is a series of modal signals with specific frequency characteristics and different frequency components reflecting the original current data of the inverter; wherein, In the variational mode decomposition method, each intrinsic mode function IMF needs to be defined as an amplitude modulation frequency modulation function u k (t) = A k (t)cos(φ k (t)) to capture the intrinsic mode of the current signal, where u k (t) is represented as the kth IMF obtained by decomposition. At the same time, it is necessary to apply Hilbert transform to each decomposed IMF to decompose the sequence data in the intrinsic mode of each current signal according to the following formula, so as to further analyze the frequency characteristics of each mode: Wherein, πt represents the phase term related to time; j represents the imaginary unit; δ(t) represents the function of time t, which is related to the phase of the signal.

6. The inverter status early warning method according to claim 5, characterized in that: For each of the modes, it is also necessary to use an exponential tuning operator to modulate the spectrum of the mode to baseband by the following formula to simplify the spectral characteristics of the modal signal and make it easier to process and analyze: The signal is demodulated by Gaussian smoothing technology to achieve bandwidth estimation, and based on the historical time series data I(t) of inverter current changes, combined with the mode set {u k } and the center frequency set ω k , construct a constrained variational problem: Where I(t) represents the historical time series data of inverter current change; the mode set {u k }={u1,u2…u N };ω k Represented as u k The kth center frequency, ω k ={ω1,ω2…ω N }; * represents the convolution operation; δ(t) represents the impulse function.

7. The inverter status early warning method according to claim 6, characterized in that: The Lagrangian function and the quadratic penalty term are introduced to transform the constrained variational problem into an unconstrained problem to simplify the problem-solving process. The process is as follows: First, define the Lagrangian function of the original constrained variational problem: In the formula, L(u k ,ω k ,λ) is represented as Lagrangian function, λ is represented as Lagrangian multiplier; α is represented as penalty term coefficient; It is expressed as a complex exponential term, which is used to represent the phase or frequency characteristics of a signal, where j is an imaginary unit; Then, a quadratic penalty term is added to the Lagrangian function; Next, the penalty term coefficient is adjusted to ensure that the solution satisfies the original constraint; Finally, the numerical optimization method is used to optimize the unconstrained objective function in the modified Lagrangian function to find the optimal solution to the problem.

8. The inverter status early warning method according to claim 7, characterized in that: When using numerical optimization methods to optimize the unconstrained objective function in the modified Lagrangian function, it is also necessary to optimize the modal function u k and the center frequency ω k The functional is updated in both directions to find the parameter value that minimizes the unconstrained objective function, which is directly related to the objective of the original constrained variational problem. The process is as follows: First, for all ω ≥ 0, we have: Where n is the number of iterations; (^) is the Fourier transform; Second, when the convergence criterion is satisfied When , the iteration is completed.

9. The inverter status early warning method according to claim 1, characterized in that: After the variational mode decomposition method is used to decompose the modal signal and obtain all the intrinsic mode functions IMF, it is also necessary to calculate the residual value of the original modal signal and the sum of all the intrinsic mode functions IMF to evaluate the quality of the decomposition, where: If the residual is smaller, it indicates that the intrinsic mode function IMF maximizes the approximation of the original modal signal; otherwise, it is necessary to readjust the parameters of the variational mode decomposition method and decompose again.

10. The inverter status early warning method according to claim 1, characterized in that: In S2, when the intrinsic mode function and weather data are used as training data and input into the recurrent neural network model for training, the following operations are performed: Normalizing the continuous moment input values ​​obtained from the intrinsic mode function and weather data, and constructing a corresponding time series data set, which is input into a recurrent neural network model; Perform forward propagation calculations to calculate the hidden state and output at the current moment based on the recurrent neural network model using the input values ​​at the current moment and the previous moment; Define a loss function to measure prediction accuracy and execute a back-propagation algorithm to calculate the gradient of the loss function with respect to the recurrent neural network model parameters to update the model parameters; Repeat the forward propagation, loss calculation, back propagation and parameter update steps through iterative training until the performance of the recurrent neural network model is stable on the validation set or reaches the set number of iterations; Meanwhile, when training the recurrent neural network model, the time series data set is divided into a training set and a test set, and the recurrent neural network model is trained using the training set; The trained model is used to predict the test set data, and the mean square error, root mean square error, and mean absolute error indicators are used to evaluate the prediction accuracy of the recurrent neural network model. The calculation formula is as follows: In the formula, N represents the number of samples; yi represents the actual value; Expressed as predicted value.

Images