Photovoltaic power prediction method fusing multi-source data

Abstract

The invention discloses a photovoltaic power prediction method fusing multi-source data, belongs to the technical field of photovoltaic power generation, and aims to solve the problems that an existing photovoltaic power prediction method depends on limited data sources and input features, so that a model is insufficient in prediction accuracy in a changeable weather environment, only basic features are utilized, a feature processing method is simple, and the prediction accuracy is low. And the prediction performance of the model is not ideal. The method comprises the following steps: acquiring external meteorological data corresponding to each station according to latitude and longitude information; acquiring historical operation data and local meteorological data of each station; interpolating the external meteorological data, and processing local meteorological data to construct a new meteorological input variable; new features are constructed by the original features through historical translation difference and rolling window statistics; constructing a second-order difference feature based on the historical translation difference feature and the rolling window statistical feature; and a training set and a test set are constructed, so that the prediction accuracy in a changeable weather environment is improved, and the prediction precision of the photovoltaic power is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to a photovoltaic power prediction method integrating multi-source data. Background Art

[0002] The output power of photovoltaic power stations is affected by factors such as weather changes. As the penetration rate of photovoltaic power generation continues to increase in new power systems, its significant randomness and volatility pose challenges to the real-time dispatch and economic operation of the power grid. In order to ensure the balance of power supply and demand, reduce dispatch costs and avoid economic losses, the power grid has strict assessment standards for the power prediction accuracy of photovoltaic power stations.

[0003] The power grid assesses the photovoltaic power forecast reporting rate and accuracy of photovoltaic stations through the "Implementation Rules for Power Grid Operation Management" and "Implementation Rules for Power Auxiliary Service Management" formulated by the National Energy Administration. Research on improving the prediction accuracy of photovoltaic station power prediction methods has important practical significance for promoting the further development of photovoltaic power generation technology and optimizing the stable operation of new power systems.

[0004] Traditional photovoltaic power prediction methods mainly include prediction techniques based on physical models and statistical models. Physical models usually rely on the physical characteristics of photovoltaic power generation systems and meteorological parameters to build models, but as the system operating parameters change, the prediction accuracy of physical models may gradually decrease; statistical models use historical data and statistical methods to predict photovoltaic power. Although they perform well in dealing with linear relationships, they may not be effective when faced with complex nonlinear relationships. Photovoltaic power is affected by a variety of meteorological factors, which have complex nonlinear relationships. The above-mentioned traditional photovoltaic power prediction methods often find it difficult to capture these complex relationships, resulting in less than ideal prediction results.

[0005] The Chinese patent document CN114330923A discloses a "photovoltaic power generation prediction method based on public meteorological data", which includes: using an API call interface to obtain hourly historical meteorological data and daily meteorological data of the district or county where the photovoltaic power station is located for no less than 100 days of history and the next 24 hours from a public meteorological data provider; obtaining 5-minute historical power generation data of no less than 100 days of the photovoltaic power station from the grid-connected inverter of the photovoltaic power station or the management software of the photovoltaic power station; processing the meteorological data and the historical power generation data to align the meteorological data and the historical power generation data; selecting data from 45 days ago to 15 days ago as a training set, data from 15 days ago to the present as a validation set, and data from the next 24 hours as a test set, respectively extracting features of the photovoltaic power generation data as input data for the next prediction algorithm; using the feature data extracted from the historical meteorological data and the historical power generation data to train a regression prediction model, and using the validation set to screen out the optimal parameters and the optimal model; using the trained regression prediction model to predict the photovoltaic power generation value for the next 24 hours to obtain the final prediction result.

[0006] Although the above-mentioned photovoltaic power prediction method introduces a machine learning algorithm to capture complex nonlinear relationships, it relies on limited data sources and input features when constructing a power prediction model, and fails to introduce and fully utilize other available meteorological information, resulting in insufficient prediction accuracy of the prediction model under changeable weather conditions. Moreover, it only uses basic features, and the feature processing method is relatively simple, resulting in unsatisfactory prediction performance of the prediction model. Summary of the invention

[0007] The present invention solves the problems that the photovoltaic power prediction method in the prior art relies on limited data sources and input features, resulting in insufficient prediction accuracy of the prediction model under changeable weather conditions, and that it only uses basic features and the feature processing method is relatively simple, resulting in unsatisfactory prediction performance of the prediction model. A photovoltaic power prediction method that integrates multi-source data is provided, which adopts diversified data sources to improve the prediction accuracy under changeable weather conditions, and adopts multiple feature construction methods to improve the prediction accuracy of photovoltaic power.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] A photovoltaic power prediction method integrating multi-source data comprises the following steps:

[0010] S1: Obtain the external meteorological data corresponding to each station according to the latitude and longitude information;

[0011] S2: Obtain historical operation data and local meteorological data of each station;

[0012] S3: interpolating the external meteorological data and processing the local meteorological data to construct new meteorological input variables;

[0013] S4: Construct new features from the original features through historical translation difference and rolling window statistics;

[0014] S5: construct second-order difference features based on historical translation difference features and rolling window statistical features respectively;

[0015] S6: Build training and test sets.

[0016] The present invention obtains external meteorological data, local meteorological data and historical operation data of each station in a distributed manner, and the data sources are diversified, which helps to improve the prediction accuracy in a changeable weather environment.

[0017] Preferably, the step S1 obtains the external meteorological data through an application program interface, the external meteorological data is historical data, and the temperature, direct irradiance and diffuse irradiance in the external meteorological data constitute external meteorological data variables. Obtaining the external meteorological data through an application program interface means obtaining meteorological data from a public meteorological data source by calling an application program interface, i.e., an API; the time interval of the obtained external meteorological data is set to 1 hour.

[0018] Preferably, the local meteorological data is historical data, and the local meteorological data includes water cycle data, atmospheric thermal data and atmospheric dynamic data. The relative humidity, precipitation and cloud cover in the water cycle data constitute water cycle data variables, the temperature, air pressure and irradiation intensity in the atmospheric thermal data constitute atmospheric thermal data variables, the 10-meter wind speed and 10-meter wind direction and 100-meter wind speed and 100-meter wind direction in the atmospheric dynamic data constitute atmospheric dynamic data variables, and the water cycle data variables, the atmospheric thermal data variables and the atmospheric dynamic data variables constitute local meteorological data variables. Local meteorological data refers to historical meteorological data stored in the database of each photovoltaic station, and the time interval of local meteorological data is set to 15 minutes.

[0019] Preferably, step S3 includes performing cubic spline interpolation on the temperature data in the external meteorological data at intervals of 15 minutes, and performing linear interpolation on the direct irradiance data and the scattered irradiance data in the external meteorological data at intervals of 15 minutes. The new meteorological input variables in step S3 are generated by algebraic operations on the local meteorological data variables, and the original features in step S4 include the external meteorological data variables, the local meteorological data variables, and the new meteorological input variables. The new meteorological input variables generated by algebraic operations on the local meteorological data variables include relative humidity / temperature, irradiance intensity / temperature, irradiance intensity / cloudiness, and 100-meter wind speed / 10-meter wind speed.

[0020] Preferably, the step S4 constructs new features from the original features by historical translation difference, comprising the following steps:

[0021] S4.1a: Setting a translation period sequence, wherein the translation period sequence includes a short-term translation period sequence and a longer-term translation period sequence, and the translation period sequences are all finite sequences;

[0022] S4.2a: For each translation period T in the translation period sequence, the original feature is translated backward by T time steps to generate a translation feature;

[0023] S4.3a: Calculate the difference between the translation feature and the original feature to generate a first-order difference feature;

[0024] S4.4a: Calculate the mean of the translation features obtained by translating the original features backward by 24 hours, 48 ​​hours, and 72 hours, respectively, to generate mean feature A, calculate 3 times the mean feature A and the mean of the translation features obtained by translating the original features backward by 96 hours and 120 hours, respectively, to generate mean feature B.

[0025] The short-term translation cycle sequence is set to: 15 minutes, 30 minutes, 45 minutes, 60 minutes, 75 minutes, 90 minutes, 105 minutes, 120 minutes, 240 minutes, 360 minutes, 480 minutes, 600 minutes; the longer-term translation cycle sequence is set to: 24 hours, 48 ​​hours, 72 hours, 96 hours, 120 hours.

[0026] Preferably, the step S4 constructs new features from the original features through rolling window statistics, comprising the following steps:

[0027] S4.1b: Setting a window size sequence, wherein the window size sequence is a finite arithmetic sequence;

[0028] S4.2b: For each window size w in the window size sequence, the average value, maximum value and minimum value of the original feature in the rolling window are calculated to generate a rolling window average value feature, a rolling window maximum value feature and a rolling window minimum value feature.

[0029] The window size sequence is set to: 4, 8, 12, 16, 20, 24.

[0030] Preferably, the width of the rolling window is 1, the height is w, and the rolling window contains w data within the previous w / 4 hours. In combination with the setting of the above window size sequence, step S4.2b generates rolling window average value features, rolling window maximum value features, and rolling window minimum value features under different window sizes of 1 hour, 2 hours, 3 hours, 4 hours, 5 hours, and 6 hours.

[0031] Preferably, the historical translation differential feature includes the translation feature and the first-order differential feature and the mean feature A and the mean feature B, and the first-order differential feature is subjected to a differential operation to generate a historical translation second-order differential feature. Specifically, for each translation period T, the feature obtained by translating the first-order differential feature backward by T time steps is subtracted from the first-order differential feature to generate a historical translation second-order differential feature.

[0032] Preferably, the rolling window statistical features include the rolling window average feature, the rolling window maximum feature, and the rolling window minimum feature, and the features obtained by subtracting the original features from the rolling window average feature, the rolling window maximum feature, and the rolling window minimum feature constitute the rolling window second-order differential features. The rolling window second-order differential features include three new features obtained by the above-mentioned subtraction operation.

[0033] Preferably, the characteristic variables of the training set and the test set include not only the historical operation data variables and the original features composed of the actual power generation of the photovoltaic units in the historical operation data, but also the historical translation difference features and the historical translation second-order difference features constructed based on them and the rolling window statistical features and the rolling window second-order difference features constructed based on them. By adopting a variety of feature construction methods, the dynamic information and trend changes in the historical meteorological data are effectively extracted, which helps to improve the prediction accuracy of photovoltaic power.

[0034] Compared with the prior art, the present invention has the following advantages: (1) the external meteorological data, local meteorological data and historical operation data of each station are obtained in a distributed manner, which can more comprehensively capture the various meteorological factors affecting photovoltaic power generation; (2) the data source is diversified, which helps to improve the prediction accuracy of the model in a changeable weather environment; (3) a variety of feature construction methods are used to effectively extract the dynamic information and trend changes in historical meteorological data, thereby enhancing the sensitivity of the model to changes in meteorological data; (4) it helps to improve the prediction accuracy of the photovoltaic power of the station and improve the operating efficiency of the photovoltaic station. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is the overall flow chart of the present invention.

[0036] Figure 2 It is a flow chart of constructing new features through historical translation difference of the present invention.

[0037] Figure 3 It is a flow chart of constructing new features through rolling window statistics of the present invention. DETAILED DESCRIPTION

[0038] The technical solution of the present invention is further described in detail below through specific embodiments and in conjunction with the accompanying drawings.

[0039] Example:

[0040] like Figures 1 to 3 The photovoltaic power prediction method integrating multi-source data shown solves the problem that the photovoltaic power prediction method in the prior art relies on limited data sources and input features, resulting in insufficient prediction accuracy of the prediction model under a changeable weather environment, and that it only uses basic features and the feature processing method is relatively simple, resulting in the prediction performance of the prediction model being not ideal. This embodiment adopts a variety of data sources to improve the prediction accuracy of photovoltaic power under a changeable weather environment, and adopts a variety of feature construction methods to improve the prediction accuracy of photovoltaic power. First, the longitude and latitude information of each photovoltaic station is obtained, and the external meteorological data corresponding to each photovoltaic station is obtained based on this information. The longitude and latitude information of each photovoltaic station is obtained through an industry database. The longitude and latitude information is the longitude and latitude coordinates, which can accurately locate the specific location of the photovoltaic station. The external meteorological data is obtained through an application program interface. Specifically, the external meteorological data is obtained from a public meteorological data source by calling an application program interface, i.e., an API. This method can realize automatic data acquisition and improve data acquisition efficiency. Then, the external meteorological data is detected for anomalies. If there is an anomaly, the application program interface is called again for acquisition. The acquired external meteorological data is the previous historical data, and the time interval of the external meteorological data is set to 1 hour. The external meteorological data variables are composed of temperature, direct irradiance and diffuse irradiance in the external meteorological data, and the external meteorological data variables are key meteorological factors that affect photovoltaic power generation.

[0041] Direct irradiance and diffuse irradiance together constitute the total irradiance, which is the total solar radiation energy vertically projected onto a unit area per unit time. Under different weather conditions and atmospheric conditions, the ratio of direct irradiance to diffuse irradiance will vary. Direct irradiance refers to the part of solar radiation that directly reaches the earth's surface from the sun. This part of light is not scattered by particles or molecules in the atmosphere. Direct irradiance accounts for a large proportion of total irradiance in clear weather. Diffuse irradiance refers to the part of solar radiation that reaches the ground due to scattering caused by gas molecules, dust particles, etc. in the atmosphere. In cloudy or foggy conditions, the contribution of diffuse irradiance to total irradiance is more significant. The level of irradiance will directly affect the size of photovoltaic power generation. Generally speaking, under the same conditions, such as the same photovoltaic module temperature and the same spectral distribution, the higher the irradiance, the more light energy the photovoltaic panel absorbs, and therefore the greater the photovoltaic power generation.

[0042] Then, the historical operation data and local meteorological data corresponding to each photovoltaic station are obtained, and the obtained historical operation data and local meteorological data are detected for abnormalities. If there are abnormalities, abnormal data processing is performed. The historical operation data refers to the actual power generation data of the photovoltaic units stored in the database of each photovoltaic station, and the time interval of the historical operation data is set to 15 minutes; the local meteorological data refers to the historical meteorological data stored in the database of each photovoltaic station. These data are collected and stored in real time by the meteorological monitoring equipment in the station. The local meteorological data obtained are also the previous historical data, and the time interval of the local meteorological data is set to 15 minutes. The local meteorological data covers the data of water cycle, atmospheric heat and atmospheric dynamics. Among them, the water cycle data variables are composed of variables such as relative humidity, precipitation and cloud cover; the atmospheric thermal data variables include variables such as temperature, air pressure and irradiation intensity; and the atmospheric dynamic data variables include variables such as 10-meter wind speed, 10-meter wind direction, 100-meter wind speed and 100-meter wind direction. These water cycle data variables, atmospheric thermal data variables and atmospheric dynamic data variables together constitute the complete local meteorological data variables.

[0043] The water cycle data variables, atmospheric thermal data variables and atmospheric dynamic data variables will directly or indirectly affect the photovoltaic power generation of the photovoltaic station, so they are of great significance for photovoltaic power prediction. Specifically, among the water cycle data variables, relative humidity affects atmospheric transparency, precipitation affects the transmittance of solar radiation, and cloud cover directly determines whether solar radiation can reach the ground; among the atmospheric thermal data variables, temperature affects the power generation efficiency of photovoltaic modules, air pressure reflects the atmospheric conditions, and irradiation intensity directly determines the photovoltaic power generation; among the atmospheric dynamic data variables, wind speed and wind direction not only affect the surface temperature of the module, but also affect the degree of dust accumulation, thereby affecting the power generation efficiency. By comprehensively considering these local meteorological data variables, the environmental factors affecting photovoltaic power generation can be more comprehensively reflected, thereby improving the accuracy and reliability of photovoltaic power generation prediction.

[0044] This embodiment obtains the historical operation data of each station in a distributed manner, local meteorological data including water cycle data, atmospheric thermal data, and atmospheric dynamic data, and external meteorological data including temperature data, direct irradiance data, and scattered irradiance data, so that various meteorological factors affecting photovoltaic power generation can be captured more comprehensively. The data source is diversified, which helps to improve the prediction accuracy of the model in a changeable weather environment, such as complex meteorological conditions such as cloudy and sunny weather, rainy and snowy weather, and sandstorm weather. This fusion and utilization of multi-source data can not only avoid the information loss or deviation that may be caused by a single data source, but also improve the robustness of the prediction model through the complementarity of data of different dimensions.

[0045] In order to improve the time resolution of the external meteorological data and keep it consistent with the time interval of other data, it is necessary to perform appropriate interpolation processing on the external meteorological data. Therefore, the next step is to interpolate the external meteorological data, and perform cubic spline interpolation processing on the temperature data in the external meteorological data, with a time interval of 15 minutes; perform linear interpolation processing on the direct irradiance data and scattered irradiance data in the external meteorological data, and the time interval is also 15 minutes. The time interval of the temperature data in the external meteorological data is set to 1 hour, and 3 new data points are inserted between every two adjacent hourly data points by the method of cubic spline interpolation, thereby subdividing the original 1-hour time interval into a 15-minute time interval; similarly, the time interval of the direct irradiance data and scattered irradiance data in the external meteorological data is also set to 1 hour, and 3 new data points are inserted between every two adjacent hourly data points by the method of linear interpolation, thereby subdividing the original 1-hour time interval into a 15-minute time interval. Since the time interval of historical operation data and the time interval of local meteorological data are both set to 15 minutes, the above-mentioned interpolation processing of the external meteorological data enables it to obtain a more detailed time resolution, providing a richer and more accurate data basis for subsequent photovoltaic power prediction, and making the time interval setting of external meteorological data consistent with the time interval setting of historical operation data and local meteorological data, so as to facilitate data processing and analysis in subsequent processes.

[0046] Cubic spline interpolation uses low-order polynomials to approximate data within a small interval, and uses the connection conditions of these polynomials to ensure the smoothness of the entire curve. The temperature changes smoothly over time. Cubic spline interpolation can capture certain nonlinear changes while maintaining smoothness, and avoid the oscillation phenomenon that may be caused by high-order polynomial interpolation. Therefore, cubic spline interpolation can better simulate the temperature change process in nature, neither too smooth nor producing unreasonable fluctuations. The changes in direct irradiance and diffuse irradiance in a short time are usually approximately linear, and there is no need to consider complex nonlinear relationships. Linear interpolation can well reflect the gradual change process of irradiance over time in a short time, avoiding the overfitting problem that may be caused by more complex interpolation methods. Moreover, linear interpolation calculations are simple and fast, and can efficiently process a large amount of irradiance data. Therefore, using linear interpolation to process irradiance data can not only ensure computational efficiency, but also better reflect the changing characteristics of actual physical phenomena. The above cubic spline interpolation operations on temperature data and linear interpolation operations on irradiance data are implemented using numerical calculation software or written code.

[0047] At the same time, local meteorological data is processed to construct new meteorological input variables. The new meteorological input variables are generated by performing reasonable algebraic operations on local meteorological data variables. The generated new meteorological input variables include relative humidity / temperature, irradiation intensity / temperature, irradiation intensity / cloud cover, and 100-meter wind speed / 10-meter wind speed. Relative humidity / temperature reflects the relationship between water vapor content and temperature in the air, which can help evaluate the stability of the atmosphere and the potential for cloud formation, and has an impact on photovoltaic power generation efficiency. Irradiation intensity / temperature reflects the irradiation intensity at unit temperature, which helps to evaluate the working efficiency of photovoltaic panels, because the increase in temperature will reduce the photoelectric conversion efficiency, which can help predict the actual output power of photovoltaic panels. Irradiation intensity / cloud cover indicates the degree of cloud cover blocking solar radiation, which helps to evaluate the actual radiation intensity reaching the ground, and is very helpful for predicting photovoltaic power generation in the short term. 100-meter wind speed / 10-meter wind speed reflects the wind speed ratio at different heights, which helps to evaluate the atmospheric stability and turbulence intensity, and has an impact on the heat dissipation conditions around the photovoltaic panels, which in turn affects the power generation efficiency. These new meteorological input variables create more informative and predictive features, helping to improve the accuracy of PV power forecasts.

[0048] Next, the historical shift difference and rolling window statistics methods are used to construct new features based on the original features. The original features here refer to a set of variables consisting of external meteorological data variables, local meteorological data variables and new meteorological input variables.

[0049] First, the historical translation difference method is used to construct new features based on the original features. The specific process of constructing new features is as follows: Figure 2As shown. Set a translation cycle sequence, which is divided into two categories, including a short-term translation cycle sequence and a longer-term translation cycle sequence. The above-mentioned translation cycle sequences are all required to be finite sequences. Determine whether the set translation cycle sequence meets the requirements. If not, reset the translation cycle sequence. The short-term translation cycle sequence is set to: 15 minutes, 30 minutes, 45 minutes, 60 minutes, 75 minutes, 90 minutes, 105 minutes, 120 minutes, 240 minutes, 360 minutes, 480 minutes, 600 minutes; the longer-term translation cycle sequence is set to: 24 hours, 48 ​​hours, 72 hours, 96 hours, 120 hours. Traverse the translation cycle sequence, and for each translation period T therein, perform a translation operation on the original feature. Specifically, translate it backward by T time steps to generate a translation feature. Then, by calculating the difference between the translation feature and the original feature, a first-order differential feature is generated. Calculate the mean of the translation features of the specified longer-term translation period to capture the trend of the longer period and construct the mean feature. Specifically, for the original features, perform the operation of backward translation of 24 hours, 48 ​​hours and 72 hours respectively to obtain the corresponding translation features. Calculate the mean of these translation features and define it as mean feature A. Multiply the value of mean feature A by 3 and calculate its mean with another set of translation features. Another set of translation features is obtained by translating the original features backward by 96 hours and 120 hours respectively. The final calculation result is defined as mean feature B.

[0050] Then, the rolling window statistics method is used to construct new features based on the original features. The specific process of constructing new features is as follows: Figure 3 As shown. Set the window size sequence, which is required to be a finite and equidistant sequence. Determine whether the set window size sequence meets the requirements. If not, reset the window size sequence. The window size sequence is set to: 4, 8, 12, 16, 20, 24. Traverse the window size sequence, and for each window size w therein, generate the rolling window average feature, rolling window maximum feature, and rolling window minimum feature by calculating the average, maximum, and minimum values ​​of the original features in the corresponding rolling window. The width of the rolling window is fixedly set to 1, and its height is set to w. The rolling window covers a total of w data in the previous w / 4 hours. Combined with the above-mentioned window size sequence settings, what is currently generated are the rolling window average feature, rolling window maximum feature, and rolling window minimum feature for different window sizes of 1 hour, 2 hours, 3 hours, 4 hours, 5 hours, and 6 hours forward.

[0051] After completing the above new features based on the original features by using the historical translation difference and rolling window statistics methods, the second-order difference features are further constructed based on the historical translation difference features and rolling window statistics features.

[0052] First, the corresponding second-order differential features are constructed based on the historical translation differential features. The historical translation differential features are composed of translation features, first-order differential features, mean features A and mean features B. The first-order differential features are subjected to a differential operation. Specifically, for each translation period T, the first-order differential features are shifted backward by T time steps and the first-order differential features are subtracted to obtain the historical translation second-order differential features.

[0053] Then, the corresponding second-order differential features are constructed based on the rolling window statistical features. The rolling window statistical features include the above-mentioned rolling window average feature, rolling window maximum feature and rolling window minimum feature. The original features are subtracted from the above-mentioned three features included in the rolling window statistical features to obtain the rolling window second-order differential features. The rolling window second-order differential features include three new features obtained by the above-mentioned subtraction operation.

[0054] After completing the construction of the above series of feature variables, the construction of the training set and the test set is carried out. The feature variables of the training set and the test set include historical operation data variables, original features, historical translation difference features, historical translation second-order difference features, rolling window statistical features, and rolling window second-order difference features, among which the historical operation data variables are the actual power generation of the photovoltaic units in the historical operation data. A variety of feature construction methods are used to effectively extract the dynamic information and trend changes in the historical meteorological data, enhance the sensitivity of the model to the changes in meteorological data, and help improve the prediction accuracy of the photovoltaic power of the station and improve the operation efficiency of the photovoltaic station. The various types of historical data obtained in this embodiment are all data from the past three years. After the above interpolation processing, the data volume is in the order of 100,000. The data volume of the training set and the test set is divided into a ratio of 8:2. Specifically, 80% of the data in the historical data that is relatively early is divided into the training set, and 20% of the data in the historical data that is relatively late is divided into the test set.

[0055] Next, the constructed training set and test set are used to train and evaluate the machine learning model. In this embodiment, the machine learning model used for photovoltaic power prediction adopts the LightGBM algorithm model. The LightGBM algorithm is an efficient and scalable machine learning algorithm based on the gradient boosting decision tree. It proposes a series of optimization strategies for data compression and decision tree modeling processes based on the original accuracy improvement optimization strategy, so that it can achieve both high computing efficiency and high computing accuracy, and solves the problem of low computing efficiency of the gradient boosting decision tree algorithm framework when processing massive data.

[0056] After completing the configuration of important model parameters such as boosting type, learning rate, number and depth of trees, the LightGBM algorithm model is trained using the training set. This embodiment adopts parallel training and uses early stopping to prevent overfitting during training. After completing the model training, the trained model is evaluated using the test set. The model evaluation indicators used in this embodiment include mean square error, root mean square error, and mean absolute error. In this embodiment, the trained LightGBM algorithm model is used for ultra-short-term photovoltaic power prediction, such as photovoltaic power prediction within the next few hours. The algorithm has significant advantages in processing large-scale data and capturing complex nonlinear relationships with its efficient processing speed and excellent prediction performance, which helps this embodiment to achieve high-precision prediction of photovoltaic power.

[0057] The trained LightGBM algorithm model can also update parameters through subsequent continuous learning to gradually improve its prediction performance. As time goes by, a certain amount of new historical data can be obtained in the same way, including external meteorological data, local meteorological data and historical operation data of each station. The new historical data is processed in the same way and used for further training of the trained LightGBM algorithm model. The model update method described in this embodiment, which continues to learn based on previous learning results, does not require training the model from scratch and is more efficient.

[0058] The embodiments described above are only preferred solutions of the present invention and are not intended to limit the present invention in any form. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the invention should be included in the protection scope of the present invention.

Claims

1. A photovoltaic power prediction method integrating multi-source data, characterized in that: The following steps are involved: S1: Obtain the external meteorological data corresponding to each station according to the latitude and longitude information; S2: Obtain historical operation data and local meteorological data of each station; S3: interpolating the external meteorological data and processing the local meteorological data to construct new meteorological input variables; S4: Construct new features from the original features through historical translation difference and rolling window statistics; S5: construct second-order difference features based on historical translation difference features and rolling window statistical features respectively; S6: Build training and test sets.

2. A photovoltaic power prediction method integrating multi-source data according to claim 1, characterized in that: The step S1 acquires the external meteorological data through an application program interface. The external meteorological data is historical data. The temperature, direct irradiance and scattered irradiance in the external meteorological data constitute external meteorological data variables.

3. The photovoltaic power prediction method integrating multi-source data according to claim 2 is characterized in that: The local meteorological data is historical data, and the local meteorological data includes water cycle data, atmospheric thermal data and atmospheric dynamic data. The relative humidity, precipitation and cloud cover in the water cycle data constitute water cycle data variables, the temperature, air pressure and radiation intensity in the atmospheric thermal data constitute atmospheric thermal data variables, the 10-meter wind speed and 10-meter wind direction and 100-meter wind speed and 100-meter wind direction in the atmospheric dynamic data constitute atmospheric dynamic data variables, and the water cycle data variables, the atmospheric thermal data variables and the atmospheric dynamic data variables constitute local meteorological data variables.

4. The photovoltaic power prediction method integrating multi-source data according to claim 3 is characterized in that: The step S3 includes performing cubic spline interpolation on the temperature data in the external meteorological data at time intervals of 15 minutes, and performing linear interpolation on the direct irradiance data and the scattered irradiance data in the external meteorological data at time intervals of 15 minutes. The new meteorological input variable in the step S3 is generated by the local meteorological data variable through algebraic operation, and the original features in the step S4 include the external meteorological data variable, the local meteorological data variable and the new meteorological input variable.

5. The photovoltaic power prediction method integrating multi-source data according to claim 1 is characterized in that: The step S4 constructs new features from the original features by historical translation difference, and includes the following steps: S4.1a: Setting a translation period sequence, wherein the translation period sequence includes a short-term translation period sequence and a longer-term translation period sequence, and the translation period sequences are all finite sequences; S4.2a: For each translation period T in the translation period sequence, the original feature is translated backward by T time steps to generate a translation feature; S4.3a: Calculate the difference between the translation feature and the original feature to generate a first-order difference feature; S4.4a: Calculate the mean of the translation features obtained by translating the original features backward by 24 hours, 48 ​​hours, and 72 hours, respectively, to generate mean feature A, calculate 3 times the mean feature A and the mean of the translation features obtained by translating the original features backward by 96 hours and 120 hours, respectively, to generate mean feature B.

6. The photovoltaic power prediction method integrating multi-source data according to claim 1 is characterized in that: The step S4 constructs new features from the original features through rolling window statistics, including the following steps: S4.1b: Setting a window size sequence, wherein the window size sequence is a finite arithmetic sequence; S4.2b: For each window size w in the window size sequence, the average value, maximum value and minimum value of the original feature in the rolling window are calculated to generate a rolling window average value feature, a rolling window maximum value feature and a rolling window minimum value feature.

7. The photovoltaic power prediction method integrating multi-source data according to claim 6 is characterized in that: The width of the rolling window is 1, and the height is w. The rolling window contains w data within the previous w / 4 hours.

8. The photovoltaic power prediction method integrating multi-source data according to claim 5 is characterized in that: The historical translation differential feature includes the translation feature and the first-order differential feature as well as the mean feature A and the mean feature B. The first-order differential feature is subjected to a differential operation again to generate a historical translation second-order differential feature.

9. The photovoltaic power prediction method integrating multi-source data according to claim 6, characterized in that: The rolling window statistical features include the rolling window average feature, the rolling window maximum feature and the rolling window minimum feature. The features obtained by subtracting the original features from the rolling window average feature, the rolling window maximum feature and the rolling window minimum feature respectively constitute the rolling window second-order differential features.

10. A photovoltaic power prediction method integrating multi-source data according to any one of claims 1 to 9, characterized in that: The characteristic variables of the training set and the test set include not only the historical operation data variables constituted by the actual power generation of the photovoltaic units in the historical operation data and the original features, but also the historical translation difference features and the historical translation second-order difference features constructed based on them and the rolling window statistical features and the rolling window second-order difference features constructed based on them.

Images