Multi-element periodic grid power prediction method based on Kalman filter

Abstract

The invention provides a multivariate periodic grid power prediction method based on a Kalman filter, relates to the field of new energy power generation power prediction, and mainly comprises four parts: establishment of a multivariate periodic grid, an initialization stage, a prediction stage and an updating stage. According to the method, historical power data and meteorological data of a new energy power station are used as input, a three-dimensional grid model covering power, periodic characteristics and time is constructed, a Kalman filter is used as a prediction tool, the power is continuously adjusted by combining historical data and current real-time input data, and the prediction accuracy of the new energy power station is improved. And according to the periodic characteristics and the weight between the meteorological factors, the grid data is continuously and dynamically updated in a rolling manner. According to the method, the periodic characteristics and dynamic changes of the grid data can be well captured, the method has remarkable advantages in the aspects of multivariate data fusion, noise suppression and real-time updating of the system state, the prediction precision is high, and the method is suitable for power prediction scenes with complex periodic fluctuations and variable environments.

Description

Technical Field

[0001] The present invention belongs to the field of new energy power generation power prediction, and particularly relates to a multi-periodic grid power prediction method based on a Kalman filter. Background Art

[0002] With the rapid development of renewable energy, power prediction in fields such as wind power generation and photovoltaic power generation has become a key issue in energy management and scheduling. The accuracy of power prediction directly affects the stability and efficiency of the power system. Especially when renewable energy is greatly affected by weather changes and environmental factors, traditional power prediction methods face great challenges. In order to improve the accuracy of power prediction, attempts have been made to combine multiple data sources, including historical power data, meteorological factors, and periodic characteristics, and adopt advanced prediction models to cope with complex power fluctuation patterns.

[0003] Traditional power prediction methods such as time series analysis, regression analysis, and artificial neural networks can provide relatively accurate predictions under certain conditions, but usually rely on a large amount of historical data for training and have weak adaptability to external environmental changes. Especially in the presence of noise or drastic fluctuations, the prediction accuracy is prone to decline. In addition, many methods ignore the periodic laws in power data and the dynamic changes of meteorological factors, resulting in the inability to accurately capture the internal patterns of power fluctuations.

[0004] As a recursive estimation algorithm, the Kalman filter is widely used in the state estimation and filtering processing of dynamic systems. At present, for the problem of power prediction, there have been some studies based on the Kalman filter, but most of them focus on the application of single data sources or static models. For the technical field of combining the Kalman filter with a multi-periodic grid for prediction, it is still in the stage of active exploration and development.

[0005] In order to solve the above problems, the current existing technologies are as follows:

[0006] Technical comparison with the patent CN109143147A "Power measurement method for an electric energy meter calibration device based on a Kalman filter algorithm"

[0007] Patent CN109143147A mainly relies on the Kalman filter to predict power data, without performing periodic modeling or structured processing on the input data. Its data preprocessing is limited to simple Kalman filter parameter initialization, lacking the extraction of periodic features of power data and the optimization of multivariable coupling relationships. Due to the lack of explicit modeling of the periodic features of power data (such as daily cycle, seasonal cycle), Patent CN109143147A may have insufficient prediction accuracy when facing complex periodic fluctuations. In this study, a multivariate periodic grid modeling is introduced. The periodic components of power data are extracted through Fourier transform, and the data is divided into different grid cells by combining K-means clustering, explicitly separating periodic and random fluctuations. It can accurately capture complex fluctuation patterns such as the daily cycle and seasonal cycle of power data, and through grid division and normalization processing, optimize the dynamic coupling relationship between power and meteorological factors (such as irradiance, temperature), and also have good prediction accuracy for complex environments.

[0008] Patent CN109143147A performs anti-disturbance processing on the Kalman filter by introducing the residual change rate (Jam), but its parameters (such as the process noise covariance Q and the measurement noise covariance R) are mainly selected through experimental calibration and grading, lacking a dynamic adjustment mechanism. When facing dynamic changes such as sudden weather changes or equipment aging, the parameter fixity of Patent CN109143147A may lead to a decrease in prediction accuracy. This study adopts a dynamic rolling update mechanism, feeds the output result of the Kalman filter back to the multivariate periodic grid, adjusts the periodic feature weights and noise covariance matrix of the grid cells in real time, and dynamically adjusts the parameters of the Kalman filter (such as Q, R) according to real-time observation data and historical data. Through incremental data fusion and local K-means clustering, it ensures that the grid model evolves dynamically with time and working conditions, and can still maintain a high prediction accuracy when there are sudden weather changes or equipment performance changes.

[0009] Patent CN109143147A is mainly applied to the power measurement of electricity meter calibration devices, focusing on the real-time prediction and anti-disturbance processing of lagging power values, and the prediction range is limited to the ultra-short term (from a few minutes to a few hours). Its application scenario is relatively single, not involving the long-term power prediction of new energy generation (such as photovoltaic, wind power), and its prediction ability for multiple time scales (such as short-term, long-term) is limited. This study is applicable to the power prediction of new energy generation, covering three time scales: ultra-short term, short term, and long term, and can handle power fluctuations in complex scenarios such as photovoltaic and wind power.

[0010] Technical comparison with Patent CN112234601A "A Method and System for Online Identifying Low-Frequency Oscillation Characteristic Parameters of Power Systems"

[0011] 1. Patent CN112234601A identifies the low-frequency oscillation characteristic parameters (such as oscillation frequency and damping ratio) of the power system online, which is used to prevent the power angle instability of the power system. For the low-frequency oscillation signal of 0.1 - 2.5 Hz, the power oscillation is modeled as an exponentially decaying cosine signal, and the oscillation parameters are extracted through the Extended Kalman Filter (EKF). However, this study constructs a three-dimensional grid model (power, periodic parameters, meteorological conditions) for the multi-scale periodic characteristics (such as daily cycle and seasonal cycle) of new energy power data and the meteorological coupling relationship, and dynamically adjusts the state transition matrix and noise covariance. Patent CN112234601A is oriented to power system stability monitoring and needs to process non-linear state space equations, which is limited to the low-frequency oscillation scenario. This study focuses on new energy power prediction, directly solves the problem of grid connection of wind and solar power generation, simplifies the multi-variable coupling relationship through grid modeling, reduces the computational complexity, and covers the full-scale prediction of ultra-short term (minute level), short term (day-ahead market), and long term (seasonal planning).

[0012] 2. In terms of data processing, Patent CN112234601A directly models the original power oscillation signal without periodic division or noise suppression, relying on the linearization process (Jacobian matrix) of the Extended Kalman Filter, which is sensitive to the model accuracy. In terms of dynamic update, the calculation times between sampling points are increased through the variable rate Extended Kalman Filter to improve the identification speed of low-frequency oscillation parameters, and the parameters (Q, R) are fixed, relying on experimental calibration and segmented selection, lacking self-adaptive ability.

[0013] However, this study

[0014] significantly reduces the noise interference through moving average and Fourier transform, and avoids data distribution deviation through incremental update of the grid model. Patent CN112234601A may diverge due to the accumulation of model errors during long-term operation.

[0015] Technical comparison with Patent CN113095562A "Ultra-short-term power generation prediction method and device based on Kalman filter and LSTM"

[0016] 1. Patent CN113095562A mainly relies on the combination of the Kalman filter and LSTM (Long Short-Term Memory network), corrects the meteorological data through the Kalman filter, and then uses LSTM for power prediction. Although LSTM can process time series data, this technology does not explicitly introduce a processing mechanism for periodic characteristics, especially lacking a systematic modeling of the periodic fluctuations of power data. This study introduces a multi-variable periodic grid model, which can more accurately capture the periodic characteristics of power data. Especially in long-term prediction, it can better cope with seasonal changes and periodic fluctuations.

[0017] Second, Patent CN113095562A mainly relies on the Kalman filter to correct meteorological data, and then inputs the corrected data into the LSTM for prediction. Although the Kalman filter can dynamically correct meteorological data, this technology does not clearly address the multi-dimensional data fusion problem among power, meteorological data, and periodic characteristics. In this study, through a multivariate periodic grid model, power, meteorological data, and periodic characteristics are fused in multiple dimensions, and the Kalman filter is used to dynamically adjust the weights between variables. This method can better handle the interaction between multi-dimensional data, especially in complex environments (such as sudden weather changes, seasonal variations, etc.), and can predict power more accurately. Summary of the Invention

[0018] To solve the above technical problems, the present invention proposes a multivariate periodic grid power prediction method based on the Kalman filter. This method uses the historical power data and meteorological data of a new energy power station as inputs, constructs a three-dimensional grid model including power, periodic characteristics, and time, and uses the Kalman filter for power prediction. By combining historical data with real-time input information, the Kalman filter dynamically adjusts the relative weights among power, periodic characteristics, and meteorological factors, thereby realizing continuous rolling update of grid data.

[0019] To achieve the above object, the technical solution adopted by the present invention is:

[0020] A multivariate periodic grid power prediction method based on the Kalman filter, the specific steps are as follows:

[0021] Step 1: Collect historical grid data and meteorological data of a new energy power station, and through data preprocessing, normalization processing, and periodic division, form a multivariate periodic grid;

[0022] Step 2: Input the grid data into the Kalman filter, and set parameters such as the initial state vector x0, the initial covariance matrix P0, the state transition matrix F, the observation matrix H, the process noise covariance matrix Q, and the observation noise covariance matrix R;

[0023] Step 3: Input the initialized parameters into the prediction stage of the Kalman filter, and through two core steps of state prediction and covariance prediction, predict the uncorrected power and covariance in the next stage;

[0024] Step 4: Input the preliminary power and covariance output in the prediction stage into the update stage of the Kalman filter, and by calculating the Kalman gain and judging the weight sizes of the preliminary prediction value and the actual power, further correct the prediction result.

[0025] As a further improvement of the present invention, Step 1 includes:

[0026] Step 1.1: Use the moving average method and the mean filling method to smooth the data and correct the missing values, forming tuples in the meteorology and power fields;

[0027] Step 1.2: Process each feature in the above data tuples using L1 norm normalization to form a normalized multi-variable dataset;

[0028] Step 1.3: Perform a two-dimensional discrete Fourier transform on the normalized multi-variable dataset, calculate the average amplitude of the spectrum as the threshold, and according to the threshold, mark the frequency points in the spectrum with amplitudes greater than the threshold as periodic components, otherwise mark them as non-periodic components. According to the marked results, perform periodic partitioning on the original data in the multi-variable dataset to separate the periodic components and non-periodic components;

[0029] Step 1.4: Use the K-means clustering algorithm for the feature data after periodic feature partitioning, and use the data points and their affiliated cluster labels as the nodes of the multi-variable periodic grid, thereby dividing the data into different spatial regions or clusters to form a multi-variable periodic grid structure.

[0030] As a further improvement of the present invention, in Step 2, the initial state vector x0 further includes the initial power P, the periodic parameters further include the frequency f and the amplitude A, and the relevant meteorological conditions further include the illumination amplitude G, the temperature T, and the humidity h. Among them, the frequency and amplitude in the periodic parameters refer to the frequency and amplitude of the power after Fourier transform decomposition. The specific relationship of the initial state vector x0 is as follows:

[0031] x0 = [P, f, A, G, T, V...]

[0032] Select a section from the historical power data and calculate its average value as the initial power, as shown in the following formula:

[0033]

[0034] Among them, P i represents the actual power value intercepted from a certain period of the power data, and P 初始 represents the initial power. The setting of the periodic parameters includes extracting the spectral characteristics of the meteorological data and the power data, determining the main period of the meteorology and power tuples through Fourier transform. The initialization setting of the meteorological conditions refers to taking the average value of the meteorological factors, including the illumination amplitude G, the temperature T, and the humidity h, in a certain period in the past;

[0035] The initial covariance matrix P0 is a symmetric positive definite matrix, which is used to characterize the uncertainty of each variable in the initial state vector x0 and the correlation between variables. Its diagonal elements represent the variances of the initial values of each state variable. For a dataset x1, x2...x containing n valuesn , the variance calculation formula is:

[0036]

[0037] where σ 2 represents the variance. According to the above formula, the initial power variance σ P 2 is calculated respectively, and the frequency variance σ f 2 and amplitude variance σ A 2 in the periodic characteristics, as well as the variances of each meteorological element including σ G 2 , σ T 2 , σ h 2 are used as the diagonal elements of the initial covariance matrix;

[0038] Since there are interactions among power, meteorological elements and periodic characteristics, in order to explore the mutual relationship among them and improve the prediction accuracy, covariance is introduced for research. The covariance calculation formula is:

[0039]

[0040] where X and Y represent each state variable. Taking power and light irradiance as an example, X represents power P, Y represents irradiance G, and covariance Cov(P, G) represents the mutual relationship between power and irradiance. Calculate the mutual interactions among each state variable in turn according to the above example to form the initial covariance matrix as follows:

[0041]

[0042] The state transition matrix reflects the dynamic characteristics and mutual relationships in time of state variables including power, periodic characteristics, and meteorological elements. Since there are influences among power values, periodic parameters, and meteorological factors, the specific form of the state transition matrix F is determined through regression or correlation analysis, as shown in the following formula:

[0043]

[0044] where Δt is the time interval for updating the prediction model, and c i,j is the linear regression coefficient;

[0045] The observation matrix H is a matrix used to describe the linear relationship between the observed value z k and the state variable x k . Its function is to convert the state variable x k into the observed value z k , z kIt represents the power data and meteorological data directly measured in practice. The specific conversion process is shown in the following formula:

[0046] z k = H·x k + v k

[0047] where v k is the observation noise;

[0048] In actual measurement, due to the combined influence of meteorological elements and periodic characteristics, the measured power value P can be expressed as:

[0049] P = c P ·P + c f ·f + c A ·A + c G ·G + c T ·T + c h ·V + v k

[0050] where c p ,c f ,c A ,c G ,c T ,c h are the weight coefficients of the state variables for power observation;

[0051] Since the observed value z k is a linear combination of power, periodic parameters, and meteorological factors, the observation matrix H is a weight matrix reflecting the weighting relationship. The observation matrix H can be expressed as:

[0052] H = [c P c f c A c G c T c V …]

[0053] The structural forms of the process noise covariance matrix Q and the observation noise covariance matrix R are similar to the initial covariance matrix P0. The diagonal elements of the former are the process noise variances of individual state variables in the state transition, and the non-diagonal elements are the process noise covariances between power, periodic parameters, and meteorological conditions, representing the uncertainty and correlation in the state transition process. The diagonal elements of the observation noise covariance matrix represent the error variances of each observed variable itself, and the non-diagonal elements represent the error covariances between variables.

[0054] As a further improvement of the present invention, in step 3, after the initialization phase ends, the output parameters are input into the prediction phase. The prediction phase uses the state information and covariance at the previous moment to infer the power value, cycle parameter, and meteorological data at the current moment, and is divided into state prediction and covariance prediction. The state prediction equation is as follows:

[0055]

[0056] Among them, x k-1 represents the predicted power, cycle parameter, and meteorological data and other state variables corrected by observation at the previous moment. F is the state transition matrix mentioned above. represents the preliminary prediction result without considering the observation information. This prediction process estimates the change trend of the state at the next moment based on the state of the transition matrix at the previous moment, and provides a prior estimate value for capturing the cycle law and dynamic changes;

[0057] The covariance prediction equation is as follows:

[0058]

[0059] Among them, P k-1 is the state covariance matrix at the previous moment. is the state covariance matrix at the current moment. Q is the process noise covariance matrix mentioned above. The covariance prediction uses the state transition matrix and the process covariance matrix to estimate the uncertainty of the predicted value.

[0060] As a further improvement of the present invention, in step 4, the update phase first requires the output results of the prediction phase, which include the uncorrected state variables the output covariance of the prediction phase and the observation value z k and the observation covariance matrix R. Then, by calculating the Kalman gain, it is determined which of the predicted value and the observation value at the current moment is closer to the expected value. When the prediction uncertainty is small, the Kalman gain will be more inclined to believe the prediction result. On the contrary, if the credibility of the observation value is high, the Kalman gain will be more inclined to believe the observation data. That is, the Kalman gain determines the weight distribution of the two in the state update. The calculation formula of the Kalman gain is as follows:

[0061]

[0062] Among them, K G is the Kalman gain. is the covariance of the prediction phase, and H is the observation matrix of the current phase.

[0063] After the weights are assigned by the Kalman gain, the updated state variable x kand the state covariance matrix P k , and the specific output process is as follows:

[0064]

[0065]

[0066] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0067] The present invention combines a multivariate periodic grid with a Kalman filter to achieve dynamic update of grid data. The invention can not only capture complex periodic fluctuations more effectively, but also perform power prediction efficiently in a real-time environment. Description of the Drawings

[0068] Figure 1 is a flow chart of multivariate periodic grid power prediction based on a Kalman filter;

[0069] Figure 2 is a simulation diagram of a multivariate periodic grid for power prediction based on a Kalman filter in the ultra-short-term background;

[0070] Figure 3 is a simulation diagram of a multivariate periodic grid for power prediction based on a Kalman filter in the short-term background;

[0071] Figure 4 is a simulation diagram of a multivariate periodic grid for power prediction based on a Kalman filter in the long-term background;

[0072] Figure 5 is a simulation comparison diagram of the prediction results of a multivariate periodic grid based on a Kalman filter and the measured data. Specific Embodiments

[0073] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments:

[0074] As Figure 1 shown, the process of multivariate periodic grid power prediction based on a Kalman filter is divided into four steps. The first step is multivariate periodic grid modeling, and the specific process includes data preprocessing, data normalization processing, periodic division, and finally forming a multivariate grid. In the modeling process, the methods and principles adopted in each step are as follows:

[0075] Step 1.1: Use the moving average method and the mean filling method to smooth the data and correct the missing values to form tuples in the meteorological and power fields;

[0076] Step 1.2: Use L1-norm normalization to process each feature in the above data tuples to form a normalized multivariate data set;

[0077] Step 1.3: Perform a two-dimensional discrete Fourier transform on the normalized multi-variate data set, calculate the average amplitude of the spectrum as the threshold, and according to the threshold, mark the frequency points in the spectrum with amplitudes greater than the threshold as periodic components, otherwise mark them as non-periodic components. According to the marked results, perform a periodic division on the original data in the multi-variate data set to separate the periodic components and non-periodic components;

[0078] Step 1.4: Use the K-means clustering algorithm for the feature data after the periodic feature division, and use the data points and their affiliated cluster labels as the nodes of the multi-variate periodic grid, thereby dividing the data into different spatial regions or clusters to form a multi-variate periodic grid structure.

[0079] In the second step, input the grid data into the Kalman filter. The Kalman filter performs an initial setting on the data, specifically including setting parameters such as the initial state vector x0, the initial covariance matrix P0, the state transition matrix F, the observation matrix H, the process noise covariance matrix Q, and the observation noise covariance matrix R.

[0080] In the third step, send the set parameters to the prediction stage of the Kalman filter. The prediction process is mainly divided into state prediction and covariance prediction. The state prediction equation is as follows:

[0081]

[0082] Among them, x k-1 represents the predicted power, cycle parameters, and meteorological data and other state variables corrected through observation at the previous moment. F is the state transition matrix mentioned above. represents the preliminary prediction result without considering the observation information. This prediction process estimates the change trend of the state at the next moment based on the state at the previous moment of the transition matrix. It provides a prior estimate value for capturing cycle laws and dynamic changes.

[0083] The covariance prediction equation is as follows:

[0084]

[0085] Among them, P k-1 is the state covariance matrix at the previous moment. is the state covariance matrix at the current moment. Q is the process noise covariance matrix mentioned above. The covariance prediction uses the state transition matrix and the process covariance matrix to estimate the uncertainty of the predicted value.

[0086] In the fourth step, update and correct the output result of the prediction stage. Determine the weight distribution of the predicted value and the actual data in the state update through the Kalman gain. The calculation formula of the Kalman gain is as follows:

[0087]

[0088] where K G is the Kalman gain, is the covariance in the prediction stage, and H is the observation matrix in the current stage.

[0089] After allocating weights through the Kalman gain, the updated state variable x k and the state covariance matrix P k are output. The specific output process is as follows:

[0090]

[0091] As Figure 2 shown in Figure 3 and Figure 2 illustrates the power prediction simulation process achieved by combining a multivariate periodic grid and Kalman filter technology on an ultra-short-term time scale. This simulation selects three consecutive hours of predicted power data, where each power prediction value is based on the information of the previous output power point to dynamically predict the power value at the current moment. And Figure 3 presents the power prediction situation on a short-term time scale. Similarly based on the method of multivariate periodic grid and Kalman filter, this simulation verification covers a complete 24-hour cycle.

[0092] Figure 4 illustrates the power prediction simulation process achieved by combining a multivariate periodic grid and Kalman filter technology on a long-term time scale. This simulation covers a whole year of data. In the simulation, the solar irradiance is weak and the power is low in winter (December - June), while the power increases significantly in summer (June - August) due to stronger irradiance. By dynamically adjusting seasonal changes and external disturbance factors, the model can accurately capture the seasonal fluctuations of power. This simulation dynamically adjusts power prediction based on the output of the previous moment to ensure the continuity and accuracy of the prediction, fully reflecting the impact of solar irradiance on power and demonstrating the seasonal change characteristics on a long-term time scale.

[0093] As Figure 5 shown in Figure 5 are the power prediction simulation result graphs achieved by combining a multivariate periodic grid and Kalman filter technology in the ultra-short-term, short-term, and long-term backgrounds respectively. From the prediction results, it can be seen that there is a high degree of fit between the predicted power curve and the actual data curve, meeting the prediction expectation and with high prediction accuracy.

[0094] The computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine - related instructions, microcode, firmware instructions, state - setting data, or source code or object code written in any combination of one or more programming languages, including object - oriented programming languages such as Smalltalk, C++, etc., and conventional procedural programming languages such as the "C" language or similar programming languages. The computer - readable program instructions may be executed entirely on the user's computer, partially on the user's computer, executed as a stand - alone software package, executed partially on the user's computer and partially on a remote computer, or executed entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider). In some embodiments, by using the state information of the computer - readable program instructions to customize an electronic circuit, such as a programmable logic circuit, a field - programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit can execute the computer - readable program instructions to implement various aspects of the present disclosure.

[0095] The above - mentioned are only the preferred embodiments of the present invention, and not any other form of limitation to the present invention. Any modification or equivalent change made according to the technical essence of the present invention still belongs to the scope of protection required by the present invention.

Claims

1. A multivariate periodic grid power prediction method based on Kalman filter, the specific steps are as follows, characterized in that: Step 1: Collect historical grid data and meteorological data of renewable energy power stations, and form a multivariate periodic grid through data preprocessing, normalization, and periodic division; Step 2: Input the grid data into the Kalman filter and set the parameters such as the initial state vector x0, the initial covariance matrix P0, the state transfer matrix F, the observation matrix H, the process noise covariance matrix Q and the observation noise covariance matrix R; Step 3: Input the initialized parameters into the prediction stage of the Kalman filter, and predict the uncorrected power and covariance of the next stage through the two core steps of state prediction and covariance prediction; Step 4: Input the preliminary power and covariance output in the prediction phase into the update phase of the Kalman filter. By calculating the Kalman gain, the weight of the preliminary prediction value and the actual power is determined, and the prediction result is further corrected.

2. The multivariate periodic grid power prediction method based on Kalman filter according to claim 1, characterized in that: Step 1 includes: Step 1.1: Use the moving average method and mean filling method to smooth the data and correct missing values ​​to form meteorological and power domain tuples; Step 1.2: Use L1 norm normalization to process each feature in the above data tuple to form a normalized multivariate data set; Step 1.3: Perform a two-dimensional discrete Fourier transform on the normalized multivariate data set, calculate the average amplitude of the spectrum as a threshold, and mark the frequency points in the spectrum with amplitudes greater than the threshold as periodic components based on the threshold, otherwise mark them as non-periodic components. Based on the marking results, perform periodic division on the original data in the multivariate data set to separate the periodic components from the non-periodic components. Step 1.4: The feature data after periodic feature division is clustered using the K-means clustering algorithm, and the data points and their cluster labels are used as nodes of the multivariate periodic grid, thereby dividing the data into different spatial regions or clusters to form a multivariate periodic grid structure.

3. The multivariate periodic grid power prediction method based on Kalman filter according to claim 1, characterized in that: In step 2, the initial state vector x0 includes the initial power P, the periodic parameters include the frequency f and the amplitude A, and the relevant meteorological conditions include the light amplitude G, the temperature T and the humidity h. The frequency and amplitude in the periodic parameters refer to the frequency and amplitude of the power after Fourier transform decomposition. The specific relationship of the initial state vector x0 is as follows: x0=[P,f,A,G,T,V...] Select a section from the historical power data and calculate its average value as the initial power, as shown in the following formula: Among them, P i Indicates the actual power value intercepted from a certain period of time in the power data, P 初始 It is expressed as initial power. The setting of period parameters includes extracting the spectrum characteristics of meteorological data and power data, determining the main period of meteorological and power tuples through Fourier transform, and the initial setting of meteorological conditions refers to taking the average value of meteorological factors in the past period of time, including light amplitude G, temperature T and humidity h; The initial covariance matrix P0 is a symmetric positive definite matrix used to characterize the uncertainty of each variable in the initial state vector x0 and the correlation between variables. Its diagonal elements represent the variance of the initial value of each state variable. For a data set containing n values ​​x1, x2…x n , the variance calculation formula is: where σ 2 Expressed as variance, according to the above formula, the initial power variance σ is obtained respectively P 2 , the frequency variance σ in the periodic characteristics f 2 and amplitude variance σ A 2 , and the variance of each meteorological element includes σ G 2 , σ T 2 , σ h 2 as the diagonal elements of the initial covariance matrix; Since there is interaction between power, meteorological elements and periodic characteristics, in order to explore the relationship between them and improve the prediction accuracy, covariance is introduced for research. The covariance calculation formula is: Among them, X and Y represent each state variable. Taking power and light irradiance as examples, X represents power P, Y represents irradiance G, and covariance Cov(P,G) represents the relationship between power and irradiance. According to the above example, the interaction between each state variable is calculated in turn to form the initial covariance matrix, as shown below: The state transfer matrix reflects the dynamic characteristics and mutual relationships of state variables including power, cycle characteristics, and meteorological elements over time. Due to the influence between power values, cycle parameters, and meteorological factors, the specific form of the state transfer matrix F is determined by regression or correlation analysis, as shown in the following formula: Where Δt is the time interval for updating the prediction model, c i,j is the linear regression coefficient; The observation matrix H is used to describe the observation value z k With the state variable x k The matrix of the linear relationship between the state variables x k Convert to observation z k , z k It represents the power data and meteorological data directly measured from actual conditions. The specific conversion process is shown in the following formula: z k =H·x k +v k Among them, v k is the observation noise; In actual measurement, due to the combined influence of meteorological elements and periodic characteristics, the measured power value P can be expressed as: P=c P ·P+c f ·f+c A ·A+c G ·G+c T ·T+c h ·V+v k where c p , c f , c A , c G , c T , c h is the weight coefficient of the state variable to the power observation; Since the observed value z k It is a linear combination of power, periodic parameters and meteorological factors. The observation matrix H is a weight matrix reflecting the weighted relationship. The observation matrix H can be expressed as: H=[c P c f c A c G c T c V …] The structures of the process noise covariance matrix Q and the observation noise covariance matrix R are similar to the initial covariance matrix P0. The diagonal elements of the former are the process noise variance of a single state variable in the state transition, and the non-diagonal elements are the process noise covariance between power, periodic parameters and meteorological conditions, which represent the uncertainty and correlation in the state transition process. The diagonal elements of the observation noise covariance matrix represent the error variance of each observed variable itself, and the non-diagonal elements represent the error covariance between the variables.

4. The multivariate periodic grid power prediction method based on Kalman filter according to claim 1, characterized in that: In step 3, after the initialization phase is completed, the output parameters are input into the prediction phase. The prediction phase uses the state information and covariance of the previous moment to infer the power value, cycle parameters and meteorological data of the current moment. It is divided into state prediction and covariance prediction. The state prediction equation is as follows: Among them, x k-1 It represents the state variables such as the predicted power, cycle parameters and meteorological data corrected by observation at the previous moment. F is the state transfer matrix mentioned above. It represents the preliminary prediction result without considering the observation information. The prediction process estimates the change trend of the state at the next moment based on the state of the transfer matrix at the previous moment. It provides a priori estimate to capture periodic laws and dynamic changes. The covariance prediction equation is as follows: Among them, P k-1 is the state covariance matrix of the previous moment, is the state covariance matrix at the current moment, Q is the process noise covariance matrix mentioned above, and the covariance prediction uses the state transfer matrix and the process covariance matrix to estimate the uncertainty of the predicted value.

5. The multivariate periodic grid power prediction method based on Kalman filter according to claim 1, characterized in that: In step 4, the update phase first requires the output of the prediction phase, which includes the uncorrected state variables Output covariance of the prediction phase and the observed value z k And the observation covariance matrix R, and then calculate the Kalman gain to determine which of the predicted value and the observed value at the current moment is closer to the expected value. When the uncertainty of the prediction is small, the Kalman gain will be more inclined to believe the prediction result. Conversely, if the credibility of the observed value is high, the Kalman gain will be more inclined to believe the observed data. That is, the Kalman gain determines the weight distribution of the two in the state update. The calculation formula of the Kalman gain is as follows: Where K G is the Kalman gain, is the covariance of the prediction stage, and H is the observation matrix of the current stage. After assigning weights through the Kalman gain, the updated state variable x is output k and the state covariance matrix P k The specific output process is as follows:

Images