Photovoltaic power interval prediction method based on time sequence decomposition and conformal quantile regression

Abstract

The invention relates to a photovoltaic power interval prediction method based on time sequence decomposition and conformal quantile regression, and the method comprises the steps: collecting historical power data, solar irradiance data, temperature data, relative humidity data and wind speed data of a photovoltaic power station, supplementing missing values in the photovoltaic power data through employing a linear interpolation method, and carrying out the prediction of the photovoltaic power interval. Secondly, normalizing all data by using a Z-Score standard method, dividing the processed data into a training set, a calibration set and a test set, constructing a time sequence decomposition model based on a NeuralProphet framework, obtaining a deterministic prediction model, training two conditional quantile models respectively by using a quantile regression method, and obtaining a time sequence decomposition model based on the NeuralProphet framework; the method comprises the following steps: determining a power interval upper limit and lower limit prediction model according with a preset confidence level, deploying the model in a calibration set, calculating to obtain an empirical quantile, inputting test data into the constructed model, determining a final prediction interval, and realizing interval prediction of ultra-short-term photovoltaic power.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to a photovoltaic power interval prediction method based on time series decomposition and conformal quantile regression. Background Art

[0002] As an important form of clean energy utilization, photovoltaic power generation plays a crucial role in the global energy transition process. It is of great significance for reducing carbon emissions, addressing climate change, and promoting the development of sustainable energy. However, the power output of photovoltaic power generation is comprehensively affected by various factors such as meteorological factors like solar irradiance, cloud cover changes, temperature, and the physical characteristics of photovoltaic modules themselves, showing significant volatility and uncertainty. This poses great challenges to the stable operation of the power system. Especially in power dispatching and power supply-demand balance, accurate photovoltaic power prediction has become one of the core elements to ensure the safe and efficient operation of the power system.

[0003] In the existing field of photovoltaic power prediction, traditional methods mainly include statistical methods, traditional machine learning methods, and deep learning methods. Statistical methods such as autoregressive integrated moving average are computationally simple, but are difficult to handle complex non-linear relationships. Traditional machine learning methods such as support vector machines and ensemble learning can build mapping relationships through their own learning abilities, but in the face of complex photovoltaic power change laws, their ability to capture long-term dependence relationships and highly non-linear relationships has obvious limitations. Deep learning methods such as CNN-LSTM can extract deep features, but require a large amount of training data and have high computational complexity.

[0004] In terms of prediction forms, deterministic prediction only outputs a single value and cannot show the degree of uncertainty of photovoltaic power, making it difficult to meet the power system's requirements for risk assessment and flexible dispatching. In the aspect of interval prediction, traditional methods mostly rely on specific probability distribution assumptions, such as common methods like interval expansion based on similar days and quantile regression. However, due to the heteroscedasticity of the actual photovoltaic power distribution caused by meteorological fluctuations, seasonal differences, etc., the preset probability distribution often does not match the actual situation, seriously affecting the accuracy and credibility of interval prediction, and unable to provide a reliable decision-making basis for the power system, posing great obstacles to ensuring the stable operation of the power grid and optimizing the allocation of power resources.

[0005] Current mainstream photovoltaic power prediction methods have many deficiencies when facing the complex characteristics of photovoltaic power. Traditional statistical methods and simple machine learning methods are difficult to effectively capture the non - linear and long - term dependence relationships in the photovoltaic power sequence. Although deep learning methods have certain advantages, they are computationally complex and require a large amount of data, facing challenges in practical applications. At the same time, most existing interval prediction methods are based on specific probability distribution assumptions. However, due to the heteroscedasticity of photovoltaic power affected by meteorology, seasons, and component characteristics, the prediction results under these assumptions deviate greatly from the actual situation, unable to accurately reflect the uncertainty of photovoltaic power and difficult to meet the demand for accurate prediction for the safe and stable operation of the power system. Secondly, in the construction of photovoltaic power prediction models, most models do not fully consider the inherent structural characteristics of the photovoltaic power time series. For example, the modeling of the periodic components (daily cycle and seasonal cycle) of photovoltaic power is not fine enough, and the time - dependent relationships in historical data are not fully explored, resulting in limited accuracy and reliability of the prediction model, unable to provide a reliable decision - making basis for power dispatching and affecting the efficient utilization of photovoltaic power in the power system. Summary of the Invention

[0006] The present invention precisely aims at the deficiencies existing in the prior art and provides a photovoltaic power interval prediction method based on time - series decomposition and conformal quantile regression.

[0007] To solve the above problems, the technical solutions adopted by the present invention are as follows:

[0008] A photovoltaic power interval prediction method based on time - series decomposition and conformal quantile regression, comprising the following steps:

[0009] S10: Collect historical power data, solar irradiance data, temperature data, relative humidity data, and wind speed data of a photovoltaic power station. Use linear interpolation to supplement the missing values in the photovoltaic power data, and then use the Z - Score standard method to normalize all data. Divide the processed data into a training set, a calibration set, and a test set to prepare for subsequent model training and verification.

[0010] S20: Construct a time - series decomposition model based on the NeuralProphet framework, decompose the photovoltaic power sequence into a trend component, a periodic component, and an autoregressive component, and fit them using a piece - wise linear model, a Fourier series decomposition model, and an AR - Net model respectively. Introduce daily cycle and seasonal cycle components into the periodic component to improve the modeling ability of the photovoltaic power time series.

[0011] S30: Based on the deterministic prediction model obtained in step S20, use the quantile regression method to train two conditional quantile models respectively to determine the upper and lower limit prediction models of the power interval that meet the preset confidence levels.

[0012] Then, deploy the model to the calibration set in step S10, calculate the conformal score of the samples, further analyze its conformal properties by evaluating the inconsistency of the calibration data, and then calculate the empirical quantile.

[0013] S40: Input the test data into the constructed model, and use the empirical quantile and the conditional quantile model calculated in step S30 to determine the final prediction interval, so as to realize the interval prediction of the ultra-short-term photovoltaic power.

[0014] Furthermore, in step S10, the following steps are also included:

[0015] S11: Process the data with missing values in the photovoltaic power data after linear interpolation supplementation by the Z-Score standard method. The processing formula is:

[0016]

[0017] where, x * is the standardized data; is the mean of the original data; σ is the standard deviation of the original data.

[0018] S12: Divide the photovoltaic data set into a training set D train , a calibration set D calib and a test set D test .

[0019] Furthermore, in step S20, the following steps are also included:

[0020] S21: Decompose the photovoltaic power time series. Since the photovoltaic power curve has significant periodicity and seasonal trend, based on the NeuralProphet framework, decompose the time series into multiple additive sub-series, model and predict them separately. According to the characteristics of photovoltaic output, represent the photovoltaic power time series y pv (t) as the formula:

[0021] y pv (t) = T(t) + S(t) + A(t)

[0022] where, T(t) is the trend component, reflecting the long-term change trend in the photovoltaic power time series; S(t) is the periodic component, reflecting the periodic change pattern in the photovoltaic series; A(t) is the autoregressive component, reflecting the influence of historical photovoltaic power data on the current and future photovoltaic power.

[0023] S22: Trend component prediction based on the piecewise linear model. The long-term change trend of photovoltaic power usually shows a phased linear increase or decrease. The piecewise linear model can accurately capture this change characteristic. By capturing a series of change points c1, c2,... in the photovoltaic power time series, the time series is linearly segmented, and then the change trend of each segment is linearly fitted. The trend component T(t) is expressed as the formula:

[0024] T(t) = (δ0 + I(t) T δ)t + (ρ0 + I(t) T ρ);

[0025] where, N c is the number of change points; δ0 is the initial change rate of the first linear segment; ρ0 is the initial offset; 1 ≤ j ≤ N; c j is the j-th change point; δ j is the change amount of the slope of the linear segment caused by c j ; ρ j is the offset of the corresponding linear segment, and ρ j = -c j δ j , the purpose is to ensure the continuity of T(t); I(t ≥ c j ) is an indicator function.

[0026] S23: Periodic component prediction based on Fourier series decomposition:

[0027] S231: The rotation of the Earth causes the solar irradiance to show regular changes within a day, and the revolution of the Earth and the tilt of the Earth's axis cause the solar irradiance to show periodic changes in the four seasons. Since the period p = 365.25d days of the seasonal periodic change data is non-integer, each periodic component in the photovoltaic power time series is modeled through Fourier series decomposition to overcome the influence of non-integer periods. For the periodic component with a period of p in the time series, its N p -th order Fourier term expansion expression is:

[0028]

[0029] where, a j , b j are the Fourier term coefficients with a frequency of .

[0030] S232: Since the photovoltaic power output is closely related to the solar irradiance, the photovoltaic power time series shows typical daily periodic changes and seasonal periodic changes. The daily periodic component and the seasonal periodic component in the periodic component of the photovoltaic power sequence are modeled separately.

[0031] For the daily periodic component, its periodic change amplitude is relatively stable and less affected by the long-term trend of the time series. Its expression is:

[0032]

[0033] For the seasonal periodic component, its period span is relatively long and it is greatly affected by the long-term trend of the time series. It is expressed as the product of the trend component T(t) and the Fourier terms, and the expression is:

[0034]

[0035] Therefore, the trend component S(t) of the photovoltaic power time series is expressed as the sum of the daily periodic component and the seasonal periodic component, and the expression is:

[0036] S(t) = S d (t) + S s (t).

[0037] S24: Autoregressive component prediction based on AR-Net:

[0038] S241: The solar irradiance is affected by the earth's rotation and weather changes, and its transition is relatively continuous and stable. At the same time, the physical characteristics of the photovoltaic system and its control strategy have hysteresis, resulting in the photovoltaic power output being affected by the power output at past moments, showing strong autoregression. Therefore, the autoregressive component in the photovoltaic power sequence is modeled, and its expression is.

[0039]

[0040] Among them, N a is the order of the autoregressive component; y t is the photovoltaic power output at time t; y t-1 , y t-2 , …, is the photovoltaic power output at N a historical moments, which are lag terms; ε t is the error term.

[0041] S242: Use AR-Net to predict the autoregressive component A(t) obtained in step S241: AR-Net can not only capture the linear autoregressive relationship, but its neural network structure can also improve the expression ability for non-linear time-dependent relationships. AR-Net is a feedforward neural network for autoregressive sequences, and its input is N a lag terms, and its mathematical expression is the traditional linear autoregressive model, and its expression is:

[0042]

[0043] Among them, c is a constant term; φ i is the lag term y t-i coefficient.

[0044] AR-Net uses the rectified linear unit ReLU as the activation function

[0045] σ(x) = ReLU(x) = max{0, x};

[0046] By default, AR-Net does not have a hidden layer and only contains an input layer and an output layer. Considering the complex nonlinear relationship of the photovoltaic power time series, N h hidden layers are added to AR-Net, and the outputs of each layer are as follows:

[0047] H1 = σ(W1x + b1)

[0048] H k = σ(W k H k-1 + b k )

[0049]

[0050] Among them, 2 ≤ k ≤ N h ; are the N h historical observation values received by the input layer, H i (1 ≤ i ≤ N h ) is the output of the i-th hidden layer; W i is the weight matrix; b i is the bias term; y t is the photovoltaic power prediction value; is the output layer weight matrix.

[0051] Furthermore, in step S30, the following steps are further included:

[0052] S31: Based on the deterministic prediction model constructed in step S20, train the trend, cycle, and autoregressive components obtained in step S11 on the training set D train obtained in step S12, construct a deterministic prediction model, and realize the decomposition of the photovoltaic time series.

[0053] S32: Train two conditional quantile models, set the confidence level to 1 - α, and use the quantile regression method to train the two conditional quantile models on the training set D train obtained in step S12, and their expressions are:

[0054] P{y PV ≤ Q 1-α (y PV )} = 1 - α;

[0055] Fitting the photovoltaic power time series y using a deterministic prediction model PV of the conditional quantile function, denoted as and

[0056] S33: Calculate the conformal fraction, and calculate the conformal fraction s of each sample in the calibration set D calib The expression for it is: i where s(t) is the conformal fraction of the sample at time t in D

[0057]

[0058] where s(t) is the conformal fraction of the sample at time t in D calib ; are the upper and lower limit values corresponding to the sample at time t, respectively.

[0059] S34: Calculate the empirical quantile. Set the confidence level to 1 - α. For the random variable X, its quantile Q(X) is defined as P{X ≤ Q(X)} = 1 - α, and obtain the 1 - α quantile of the empirical distribution of the non - singularity score function s. The expression for it is:

[0060]

[0061] where Q 1-α (s) is the 1 - α empirical quantile of s; n calib is the number of samples in the calibration set.

[0062] Furthermore, in step S40, the following steps are further included:

[0063] S41: Calculate the conformal quantile regression prediction interval. Based on the data obtained in step S30, the conformal quantile regression prediction interval of the test set D test data can be obtained. The expression for it is:

[0064]

[0065] where C(t) is the conformal prediction interval of the photovoltaic power at time t in the test set; Q 1-α (s) is the 1 - α empirical quantile of s.

[0066] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0067] 1. Improve the accuracy of deterministic prediction: The photovoltaic power interval prediction method based on time series decomposition and conformal quantile regression of the present invention can accurately capture the periodic information of the photovoltaic power time series by introducing the daily cycle component and the seasonal cycle component into the NeuralProphet model, greatly enhancing the learning ability of the photovoltaic power change law. At the same time, the introduced AR-Net autoregressive model effectively tracks the temporal dependence between photovoltaic time series data and successfully captures the details of photovoltaic volatility. Compared with advanced algorithms such as TimesNet and Informer, it shows obvious advantages in the deterministic photovoltaic power prediction task, effectively reducing the prediction error, significantly improving the accuracy and stability of the prediction, providing highly reliable power prediction values for the stable operation of photovoltaic power plants, and strongly guaranteeing the basic power generation plan arrangement of the power system.

[0068] 2. Enhance the reliability of interval prediction: The present invention adopts the conformal quantile regression method, abandoning the dependence on the prior probability distribution assumption of traditional interval prediction. By calculating the conformal score and empirical quantile, the function of dynamically adjusting the prediction interval according to the test data in real time is realized. Under different seasons and various weather conditions, especially when the photovoltaic output data fluctuates violently, it can still effectively cover the actual photovoltaic output data at the set confidence level. In the comparison with the traditional quantile regression algorithm, it shows excellent performance in key indicators such as prediction interval coverage rate, standardized interval width, and interval coverage width, greatly enhancing the reliability of interval prediction, providing accurate and credible uncertainty quantification information for power dispatching decisions, and effectively improving the ability of the power system to cope with the uncertainty of photovoltaic power.

[0069] Specific other beneficial effects will be described in detail in the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 is a flowchart of the photovoltaic power interval prediction method based on time series decomposition and conformal quantile regression of the present invention;

[0071] Figure 2 is a schematic diagram of the AR-Net model structure. DETAILED DESCRIPTION OF THE INVENTION

[0072] The content of the present invention will be described below in conjunction with specific implementation manners. Examples of the implementation manners are shown in the drawings, in which the same or similar reference numerals denote the same or similar components or components with the same or similar functions throughout.

[0073] The directional terms mentioned in the present invention, such as: up, down, left, right, front, back, inside, outside, front side, back side, side, etc., are only with reference to the directions in the attached drawings. The implementation manners and the directional terms used described below with reference to the attached drawings are exemplary and are only used to explain the present invention, and should not be construed as a limitation to the present invention. In addition, for the examples of various specific processes and materials provided by the present invention, those of ordinary skill in the art can realize the application of other processes and / or the use of other materials.

[0074] Please refer to Figure 1 and Figure 2 , wherein Figure 1 is a flowchart of the photovoltaic power interval prediction method based on time series decomposition and conformal quantile regression of the present invention; Figure 2 is a schematic diagram of the AR-Net model structure.

[0075] The photovoltaic power interval prediction method based on time series decomposition and conformal quantile regression includes the following steps:

[0076] S10: Collect historical power data, solar irradiance data, temperature data, relative humidity data, and wind speed data of the photovoltaic power station, supplement the missing values in the photovoltaic power data by the linear interpolation method, and then normalize all the data by the Z-Score standard method. Divide the processed data into a training set, a calibration set, and a test set to prepare for subsequent model training and verification.

[0077] S11: Perform Z-Score standard method processing on the data after linear interpolation of the missing values of the photovoltaic power data. The processing formula:

[0078]

[0079] wherein, x * is the standardized data; is the mean value of the original data; σ is the standard deviation of the original data.

[0080] S12: Divide the photovoltaic data set into a training set D train , a calibration set D calib and a test set D test .

[0081] S20: Construct a time series decomposition model based on the NeuralProphet framework, decompose the photovoltaic power sequence into a trend component, a periodic component, and an autoregressive component, and respectively fit them with a piecewise linear model, a Fourier series decomposition model, and an AR-Net model. And introduce daily cycle and seasonal cycle components into the periodic component to improve the modeling ability of the photovoltaic power time series.

[0082] S21: Photovoltaic power time series decomposition. Since the photovoltaic power curve has significant periodicity and seasonal trends, based on the NeuralProphet framework, the time series is decomposed into multiple additive subsequences, which are modeled and predicted separately. According to the characteristics of photovoltaic output, the photovoltaic power time series y pv (t) is expressed as the formula:

[0083] y pv (t) = T(t) + S(t) + A(t);

[0084] Among them, T(t) is the trend component, reflecting the long-term change trend in the photovoltaic power time series; S(t) is the periodic component, reflecting the periodic change pattern in the photovoltaic series; A(t) is the autoregressive component, reflecting the impact of historical photovoltaic power data on the current and future photovoltaic power.

[0085] S22: Prediction of the trend component based on the piecewise linear model. The long-term change trend of photovoltaic power usually shows a stagewise linear increase or decrease. The piecewise linear model can accurately capture this change characteristic. By capturing a series of change points c1, c2,... of the photovoltaic power time series, the time series is linearly segmented, and then the change trends of each segment are linearly fitted. The trend component T(t) is expressed as the formula:

[0086] T(t) = (δ0 + I(t) T δ)t + (ρ0 + I(t) T ρ);

[0087] Among them, N c is the number of change points; δ0 is the initial change rate of the first linear segment; ρ0 is the initial offset; 1 ≤ j ≤ N; c j is the j-th change point; δ j is the change amount of the slope of the linear segment caused by c j ; ρ j is the offset of the corresponding linear segment, and ρ j = -c j δ j , the purpose is to ensure the continuity of T(t); I(t ≥ c j ) is the indicator function.

[0088] S23: Prediction of the periodic component based on Fourier series decomposition:

[0089] S231: The rotation of the Earth causes regular variations in solar irradiance within a day, while the revolution of the Earth and the tilt of the Earth's axis result in periodic variations in solar irradiance over the four seasons. Since the period p = 365.25 days of the seasonal cycle change data is non-integer, Fourier series decomposition is used to model each periodic component in the photovoltaic power time series, overcoming the influence of non-integer periods. For the periodic component with a period of p in the time series, its N p -th order Fourier term expansion representation is:

[0090]

[0091] where, a j , b j are the Fourier term coefficients with a frequency of .

[0092] S232: Since the photovoltaic power output is closely related to solar irradiance, the photovoltaic power time series exhibits typical daily and seasonal periodic variations. The daily and seasonal periodic components in the periodic components of the photovoltaic power series are modeled separately.

[0093] For the daily periodic component, its periodic variation amplitude is relatively stable and less affected by the long-term trend of the time series. Its representation is:

[0094]

[0095] For the seasonal periodic component, its period span is relatively long and it is greatly affected by the long-term trend of the time series. It is expressed as the product of a trend component T(t) and a Fourier term, expressed as the formula:

[0096]

[0097] Therefore, the trend component S(t) of the photovoltaic power time series is expressed as the sum of the daily and seasonal periodic components, expressed as the formula:

[0098] S(t) = S d (t) + S s (t).

[0099] S24: Autoregressive component prediction based on AR-Net:

[0100] Solar irradiance is affected by the rotation of the Earth and weather changes, and its transition is relatively continuous and stable. At the same time, the physical characteristics of the photovoltaic system and its control strategy have hysteresis, resulting in the photovoltaic power output being affected by the power output at past moments, showing strong autoregression. Therefore, the autoregressive component in the photovoltaic power series is modeled, and its representation is:

[0101]

[0102] Among them, N a is the order of the autoregressive component; y t is the photovoltaic power output at time t; y t-1 , y t-2 , …, are the photovoltaic power outputs at N a historical moments, which are lag terms; ε t is the error term.

[0103] S242: Use AR-Net to predict the autoregressive component A(t) obtained in step S241: AR-Net can not only capture the linear autoregressive relationship, but its neural network structure can also improve the expression ability for non-linear time-dependent relationships. AR-Net is a feedforward neural network for autoregressive sequences, and its input is N a lag terms, and its mathematical representation is the traditional linear autoregressive model, and its expression is:

[0104]

[0105] Among them, c is the constant term; φ i is the lag term y t-i coefficient.

[0106] AR-Net uses the rectified linear unit ReLU as the activation function

[0107] σ(x) = ReLU(x) = max{0, x};

[0108] By default, AR-Net does not have a hidden layer and only has an input layer and an output layer. Considering the complex non-linear relationship of the photovoltaic power time series, N h hidden layers are added to AR-Net. As Figure 2 shown, the outputs of each layer are as follows:

[0109] H1 = σ(W1x + b1)

[0110] H k = σ(W k H k-1 + b k )

[0111]

[0112] Among them, 2 ≤ k ≤ N h ; are the N h historical observations received by the input layer, H i (1 ≤ i ≤ Nh ) is the output of the i-th hidden layer; W i is the weight matrix; b i is the bias term; y t is the predicted value of photovoltaic power; is the weight matrix of the output layer.

[0113] S30: Based on the deterministic prediction model obtained in step S20, using the quantile regression method, train two conditional quantile models respectively to determine the upper and lower limit prediction models of the power interval that meet the preset confidence level.

[0114] Then deploy the model to the calibration set in step S10, calculate the conformal score of the samples, further analyze its conformal properties by evaluating the inconsistency of the calibration data, and then calculate the empirical quantile.

[0115] S31: Based on the deterministic prediction model constructed in step S20, on the training set D obtained in step S12 train , train the trend, period, and autoregressive components obtained in step S11 respectively to construct a deterministic prediction model and realize the decomposition of the photovoltaic time series.

[0116] S32: Train two conditional quantile models, set the confidence level to 1-α, and use the quantile regression method to train two conditional quantile models on the training set D train respectively. Its expression is:

[0117] P{y PV ≤Q 1-α (y PV )}=1-α;

[0118] Use the deterministic prediction model to fit the conditional quantile function of the photovoltaic power time series y PV of , denoted as and

[0119] S33: Calculate the conformal score, calculate the conformal score s calib of each sample in the calibration set D i . Its expression is:

[0120]

[0121] where s(t) is the conformal score of the sample at time t in D calib ; are the upper and lower limit values corresponding to the sample at time t respectively.

[0122] S34: Calculate the empirical quantile, set the confidence level to 1-α. For the random variable X, its quantile Q(X) is defined as P{X≤Q(X)} = 1-α, and obtain the 1-α quantile of the empirical distribution of the non-singularity score function s, and its expression is:

[0123]

[0124] where Q 1-α (s) is the 1-α empirical quantile of s; n calib is the number of calibration set samples.

[0125] S40: Input the test data into the constructed model, and use the empirical quantile and the conditional quantile model calculated in step S30 to determine the final prediction interval, so as to realize the interval prediction of the ultra-short-term photovoltaic power.

[0126] S41: Calculate the conformal quantile regression prediction interval. Based on the data obtained in step S30, the conformal quantile regression prediction interval of the test set D test data can be obtained, and its expression is:

[0127]

[0128] where C(t) is the conformal prediction interval of the photovoltaic power at time t of the test set; Q 1-α (s) is the 1-α empirical quantile of s.

[0129] Improve the accuracy of deterministic prediction: The photovoltaic power interval prediction method based on time series decomposition and conformal quantile regression of the present invention can accurately capture the periodic information of the photovoltaic power time series by introducing the daily cycle component and the seasonal cycle component into the NeuralProphet model, greatly enhancing the learning ability of the change law of the photovoltaic power. At the same time, the introduced AR-Net autoregressive model effectively tracks the time dependence between the photovoltaic time series data and successfully captures the details of the photovoltaic volatility. Compared with advanced algorithms such as TimesNet and Informer, it shows obvious advantages in the deterministic photovoltaic power prediction task, effectively reducing the prediction error, significantly improving the accuracy and stability of the prediction, providing highly reliable power prediction values for the stable operation of the photovoltaic power station, and strongly guaranteeing the basic power generation plan arrangement of the power system.

[0130] Enhancing the reliability of interval prediction: The present invention adopts the conformal quantile regression method, abandoning the dependence of traditional interval prediction on the assumption of prior probability distribution. By calculating the conformal score and empirical quantile, the function of dynamically adjusting the prediction interval according to the test data is realized. Under different seasons and various weather conditions, especially when the photovoltaic output data fluctuates violently, it can still effectively cover the actual photovoltaic output data at the set confidence level. In comparison with the traditional quantile regression algorithm, it shows excellent performance in key indicators such as prediction interval coverage rate, standardized interval width, and interval coverage width, greatly enhancing the reliability of interval prediction, providing accurate and credible uncertainty quantification information for power dispatching decisions, and effectively improving the ability of the power system to cope with the uncertainty of photovoltaic power.

[0131] A method of decomposing the photovoltaic power sequence based on the NeuralProphet time series decomposition framework and effectively fitting and modeling the trend, periodic, and autoregressive components using the piecewise linear model, Fourier series decomposition model, and AR-Net model respectively, and introducing daily and seasonal periodic components into the periodic component to improve the modeling ability of the photovoltaic power time series.

[0132] Combining the techniques of conformal prediction and quantile regression, by calculating the conditional quantile, conformal score, and conformal quantile, the dynamic adjustment of the prediction interval width is realized. This method does not require a preset probability distribution, effectively handles the uncertainty of photovoltaic output, and improves the accuracy and reliability of interval prediction.

[0133] Using a specific photovoltaic power data processing process, including missing value supplementation, normalization processing, and reasonable dataset division, to provide a high-quality data basis for model training and prediction, ensuring the effectiveness of the entire prediction method.

[0134] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A photovoltaic power interval prediction method based on time series decomposition and conformal quantile regression, characterized in that: The following steps are involved: S10: Collect historical power data, solar irradiance data, temperature data, relative humidity data, and wind speed data of the photovoltaic power station, use linear interpolation to supplement the missing values ​​in the photovoltaic power data, and then use the Z-Score standard method to normalize all the data. The processed data is divided into training set, calibration set, and test set to prepare for subsequent model training and verification; S20: Construct a time series decomposition model based on the NeuralProphet framework, decompose the photovoltaic power series into trend component, periodic component, and autoregressive component, and use piecewise linear model, Fourier series decomposition model, and AR-Net model fitting respectively, and introduce daily cycle and seasonal cycle components into the periodic component to improve the modeling ability of photovoltaic power time series; S30: Based on the deterministic prediction model obtained in step S20, two conditional quantile models are trained respectively using a quantile regression method to determine a power range upper limit and a power range lower limit prediction model that meet a preset confidence level; The model is then deployed to the calibration set in step S10, the conformal score of the sample is calculated, and its conformal properties are further analyzed by evaluating the inconsistency of the calibration data, and then the empirical quantile is calculated; S40: input the test data into the constructed model, and use the empirical quantile and conditional quantile model calculated in step S30 to determine the final prediction interval, so as to achieve interval prediction of ultra-short-term photovoltaic power.

2. The photovoltaic power interval prediction method based on time series decomposition and conformal quantile regression according to claim 1 is characterized in that: In step S10, the following steps are also included: S11: For the missing values ​​of photovoltaic power data, the data supplemented by linear interpolation are processed by Z-Score standard method. The processing formula is: Among them, x * is the standardized data; is the mean of the original data; σ is the standard deviation of the original data; S12: Divide the photovoltaic data set into training set D train , calibration set D calib and the test set D test .

3. The photovoltaic power interval prediction method based on time series decomposition and conformal quantile regression according to claim 2 is characterized in that: In step S20, the following steps are also included: S21: Photovoltaic power time series decomposition. The photovoltaic power curve has significant periodicity and seasonal trend. Based on the NeuralProphet framework, the time series is decomposed into multiple additive subsequences, and modeled and predicted separately. According to the characteristics of photovoltaic output, the photovoltaic power time series y pv (t) is expressed as: y pv (t)=T(t)+S(t)+A(t); Among them, T(t) is the trend component, which reflects the long-term change trend in the photovoltaic power time series; S(t) is the periodic component, which reflects the periodic change pattern in the photovoltaic series; A(t) is the autoregressive component, which reflects the impact of historical photovoltaic power data on photovoltaic power at the current and future moments; S22: Trend component prediction based on piecewise linear model. The long-term trend of photovoltaic power usually shows a phased linear growth or decline. The piecewise linear model can accurately capture this change feature by capturing a series of change points in the photovoltaic power time series. The time series is linearly segmented, and then the changing trend of each segment is linearly fitted. The trend component T(t) is expressed as: T(t)=(δ0+I(t) T δ)t+(ρ0+I(t) T r); in, N c is the number of change points; δ0 is the initial change rate of the first linear segment; ρ0 is the initial offset; 1≤j≤N; c j is the jth change point; δ j c j The change in the slope of the linear segment caused by j is the offset of the corresponding linear segment, and ρ j =-c j δ j , the purpose is to ensure the continuity of T(t); I(t≥c j ) is the indicator function; S23: Prediction of periodic components based on Fourier series decomposition: S231: The rotation of the earth causes the solar irradiance to change regularly within a day, while the revolution of the earth and the tilt of the earth's axis cause the solar irradiance to change periodically in the four seasons. The period p = 365.25 days based on the seasonal periodic change data is non-integer. Each periodic component in the photovoltaic power time series is modeled by Fourier series decomposition to overcome the influence of non-integer periods. For the periodic component with period p in the time series, its N p The order Fourier expansion expression is: Among them, a j 、b j The frequency is The Fourier term coefficients of ; S232: Photovoltaic power output is closely related to solar irradiance. Photovoltaic power time series shows typical characteristics of daily periodic changes and seasonal periodic changes. The daily periodic component and seasonal periodic component in the periodic component of the photovoltaic power series are modeled separately; For the daily cycle component, its periodic variation amplitude is relatively stable and is less affected by the long-term trend of the time series. Its expression is: For seasonal cycle components, the cycle span is longer and is greatly affected by the long-term trend of the time series. It is expressed as the product of the trend component T(t) and the Fourier term, as follows: The trend component S(t) of the photovoltaic power time series is expressed as the sum of the daily cycle component and the seasonal cycle component, which can be expressed as: S(t)=S d (t)+S s (t); S24: Autoregressive component prediction based on AR-Net: S241: Solar irradiance is affected by the rotation of the earth and weather changes, and its transition is relatively continuous and stable. At the same time, the physical characteristics of the photovoltaic system and its control strategy have hysteresis, which causes the photovoltaic power output to be affected by the power output at the past moment, showing strong autoregression. Therefore, the autoregressive component in the photovoltaic power sequence is modeled, and its expression is: Among them, N a is the order of the autoregressive component; y t is the photovoltaic power output at time t; N a The photovoltaic power output at a historical moment is the lag term; ε t is the error term; S242: Use AR-Net to predict the autoregressive component A(t) obtained in step S241: AR-Net can not only capture linear autoregressive relationships, but its neural network structure can also improve the ability to express nonlinear time-dependent relationships. AR-Net is a feedforward neural network for autoregressive sequences, and its input is N a The mathematical representation of the lag term is the traditional linear autoregressive model, which is expressed as: Among them, c is a constant term; φ i is the lag term y t-i coefficient; AR-Net uses the rectified linear unit ReLU as the activation function σ(x)=ReLU(x)=max{0,x}; By default, AR-Net has no hidden layer, only input layer and output layer. Considering that the photovoltaic power time series has a complex nonlinear relationship, N is added to AR-Net. h There are 3 hidden layers, and the output of each layer is as follows: H1=σ(W1x+b1) H k =σ(W k H k-1 +b k ) and t =W Nh+1 HN h ; Among them, 2≤k≤N h ; is the N received by the input layer h Historical observations, H i (1≤i≤N h ) is the output of the i-th hidden layer; W i is the weight matrix; b i is the bias term; y t is the predicted value of photovoltaic power; is the output layer weight matrix.

4. The photovoltaic power interval prediction method based on time series decomposition and conformal quantile regression according to claim 3 is characterized in that: The step S30 further includes the following steps: S31: Construct a deterministic prediction model based on step S20, and use the training set D obtained in step S12 train The trend, cycle, and autoregressive components obtained in step S11 are trained separately to construct a deterministic prediction model to achieve the decomposition of the photovoltaic time series; S32: Train two conditional quantile models, set the confidence level to 1-α, and use the quantile regression method on the training set D train Two conditional quantile models are trained separately, and their expressions are: P{and PV ≤Q 1-α (and PV )}=1-α; Fitting the photovoltaic power time series y using a deterministic forecasting model PV of The conditional quantile function is denoted by and S33: Calculate the conformal score and calculate the calibration set D calib The conformal fraction s of each sample i , which is expressed as: Where s(t) is D calib The conformal fraction of the sample at time t; They are the upper and lower limit values ​​corresponding to the sample at time t respectively; S34: Calculate the empirical quantile, set the confidence level to 1-α, and for the random variable X, its quantile Q(X) is defined as P{X≤Q(X)}=1-α, and obtain the 1-α quantile of the empirical distribution of the non-singular score function s, which is expressed as: Among them, Q 1-α (s) is the 1-α empirical quantile of s; n calib The number of samples in the calibration set.

5. The photovoltaic power interval prediction method based on time series decomposition and conformal quantile regression according to claim 4 is characterized in that: The step S40 further includes the following steps: S41: Calculate the conformal quantile regression prediction interval. Based on the data obtained in step S30, the test set D can be obtained. test The conformal quantile regression prediction interval of the data is expressed as: Where C(t) is the conformal prediction interval of the photovoltaic power at time t in the test set; Q 1-α (s) is the 1-α empirical quantile of s.

Images