Multi-scale fusion prediction method for wind-solar power generation power in cold region

Abstract

The invention relates to a cold region wind-solar power generation power multi-scale fusion prediction method, and relates to the cold region wind-solar power generation power prediction technology field, and the method comprises the steps: based on actually-measured wind-solar power generation base generator set wind-solar input and power output data under different wind speeds, wind directions and illumination intensities, different regions and different time scales; carrying out characteristic analysis on the time shifting property, volatility and complementarity of the generated power of the wind-solar power generation base generator set, and constructing a power characteristic database; correcting a numerical weather forecast wind speed / irradiance error by using time sequence transmission, and establishing a wind-solar power generation power prediction model; according to the wind-solar power generation power and load combined prediction based on deep learning, regional combination is performed on wind-solar power generation power prediction models in different regions for multi-scale fusion prediction, and the data quality and the analysis accuracy are remarkably improved. In a data acquisition link, reasonable acquisition frequency setting and comprehensive collection of various data provide abundant information sources for subsequent analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of cold-region wind-solar power generation prediction, and in particular to a cold-region wind-solar power generation prediction method using multi-scale fusion. Background Art

[0002] The climate conditions in cold regions are harsh, with low temperatures and frequent snow and wind weather, which have a significant impact on the performance and operational stability of power generation equipment. At the same time, the terrain and landforms in cold regions are diverse, such as the staggered distribution of mountains, plains, and forests, which leads to uneven distribution of wind and light fields, and complex temporal and spatial variations of wind speed and irradiance. In addition, cold-region wind and solar power generation systems are usually distributed over a wide range, involving wind farms and photovoltaic power stations in multiple geographical locations. There are differences in the characteristics of wind and solar resources and power generation performance between different sites, and there are complex relationships between them such as time shift, volatility, and complementarity.

[0003] The power prediction technology of wind and solar power generation in cold regions has made significant progress in recent years. In terms of data collection, the continuous upgrading of sensor technology enables more accurate and real-time acquisition of meteorological data such as wind speed, irradiance, temperature, and the operating status data of power generation equipment. The development of communication technology also ensures the efficient transmission and aggregation of data, providing a rich data foundation for subsequent analysis and prediction.

[0004] According to a multi-model fusion adaptive new energy power prediction method and system with Chinese patent number CN 114493052 A, the method includes: obtaining multiple initial prediction models, all of which are used to predict the power of a target power generation system, and the model types of the multiple initial prediction models are different from each other; for each initial prediction model, the initial prediction model is trained using training sets of different batches to obtain multiple batch sub-models corresponding to the training sets of different batches; based on the obtained batch sub-models, model fusion processing is performed to obtain a target prediction model, and the target prediction model is used to predict the power of the target power generation system. The target prediction model obtained by this method can improve the accuracy of new energy power prediction.

[0005] Although the above-mentioned prior art has made some progress in the prediction of wind and solar power generation in cold regions, there are still some problems that need to be solved:

[0006] 1. In cold environments, sensors are easily affected by low temperatures, ice and snow, and other bad weather, resulting in abnormal values ​​and missing values ​​in data collection. Although some conventional data processing methods such as linear interpolation can fill missing values, the processing effect is often not ideal for complex continuous missing or abnormal fluctuations, which may introduce errors and affect the accuracy of subsequent analysis and prediction;

[0007] 2. The types of abnormal operation data of wind and solar power stations in cold regions are complex and diverse, and existing abnormal data recognition methods are difficult to accurately identify all types of abnormal data in some cases. For example, for accumulated abnormal data that is closely connected to normal data below the power curve (such as caused by partial equipment failure or special operating conditions), and scattered abnormal data below the power curve (such as caused by short-term power curtailment or equipment fine-tuning), traditional recognition algorithms based on fixed thresholds or simple statistical methods may make misjudgments or omissions.

[0008] In summary, wind and solar power generation power forecasting is a necessary prerequisite for increasing the proportion of wind and solar power stations connected to the grid and ensuring the safe and stable operation of a high-proportion renewable energy power system. However, a wind and solar power generation base generally includes multiple power stations with a large number of units, a wide distribution area, and diverse meteorological and geographical features. High-precision power forecasting is very difficult. In addition, the operating conditions of cold-region power grids are complex and their anti-interference ability is weak. The power forecasting of wind and solar power generation bases is one of the prerequisites for the safe and stable operation of the system. At present, the power forecasting accuracy of wind and solar power generation bases in cold-region power grids deviates greatly from actual demand. Summary of the invention

[0009] In view of the shortcomings of the prior art, the present invention provides a multi-scale fusion prediction method for wind and solar power generation in cold regions, which solves the problem that the power prediction accuracy of wind and solar power generation bases in cold-region power grids in the prior art deviates greatly from actual demand.

[0010] To achieve the above objectives, the multi-scale fusion prediction method for wind and solar power generation in cold regions of the present invention comprises the following steps:

[0011] Based on the measured wind and solar input and power output data of wind and solar power generation base generators in different regions and time scales under different wind speeds, wind directions, and light intensities, the characteristics of the time shift, volatility, and complementarity of the power generation of wind and solar power generation base generators are analyzed to build a power characteristics database;

[0012] Use time series transfer to correct the wind speed / irradiance error of numerical weather forecasts and establish a wind and solar power generation power prediction model;

[0013] Based on deep learning, the wind and solar power generation power and load joint prediction is carried out, and the wind and solar power generation prediction models in different regions are regionally combined for multi-scale fusion prediction.

[0014] Preferably, the time shift of the generated power is analyzed, and the specific steps are as follows:

[0015] Sp1. When performing the time-shift analysis, the wind speed sequences of wind turbines at different locations in the same wind farm are first collected. The sequences are collected through anemometers and meteorological towers. After the wind speed sequences of wind turbines at different locations are collected, they are preprocessed. During the preprocessing, the quality of the original wind speed data is first checked to remove abnormal values ​​and missing values. The abnormal values ​​are determined by the triple standard deviation method. When determining, if the difference between a single data point and the mean exceeds three times the standard deviation, the data point is considered to be an abnormal value. For missing values, linear interpolation is used to fill them. Linear estimation is performed based on the values ​​of adjacent data points to supplement the missing values. When linearly estimating, the wind speed sequence is set to V = [v1, v2, ..., vn], and vi is set to be a missing value, then its adjacent data point is v i-1 and v i+1 In the actual process, when the missing value is not the first and last data point of the sequence but a boundary point, only one adjacent point is taken for special processing, and the linear estimation formula is: By linear estimation formula v i Positioned at v i-1 and v i+1 The midpoint position between the ranges, when the time interval is considered in the calculation process, let the time interval be Δt, let the adjacent time points t in the time series be i-1 ,t i ,t i+1 , the corresponding wind speed values ​​are v i-1 、v i (missing values) and v i+1 ,but

[0016] Sp2. Divide the wind speed and wind direction sequences according to fixed time windows to obtain a set of wind processes. By judging the relative position of the main wind direction and the point in the wind process, select a new wind speed sequence for studying time shift, define the delay time (Δt) and acceleration factor (Δv) indicators, where the delay time indicates the time lag of the wind speed change at one point relative to another point, and the acceleration factor reflects the difference in the rate of change of the wind speed at different points. The actual distance between the two points and the actual wind conditions are constrained, and the minimum Euclidean distance between the corresponding wind processes is the optimization goal. The firefly optimization algorithm is used to solve the above indicators;

[0017] Sp3. Based on the delay time and acceleration factor of each wind process obtained, the time shift between wind speed sequences at different spatial points is analyzed, and at the same time, indicators for evaluating time shift are selected according to distance and wind power installed capacity. When the distance between wind farms exceeds 60km, it is determined that there is no acceleration factor influence between the wind speed sequences at the locations of each wind farm.

[0018] Preferably, the firefly optimization algorithm is as follows:

[0019] Assume that there are two points A and B, whose wind speed sequences are v A (t) and v B (t), t represents time. First, the delay time is preliminarily estimated by the cross-correlation function, and v is calculated. A (t) and v B The cross-correlation function of (t):

[0020]

[0021] Where τ is the time delay, N is the number of data points, and the τ value corresponding to the maximum value of the cross-correlation function is the initial estimated delay time Δt0;

[0022] Then, the initially estimated delay time is optimized, considering the actual distance d between the two points and the actual wind conditions of the average wind speed and wind direction changes as constraints, and taking the minimum Euclidean distance between the corresponding wind processes as the optimization goal. The firefly optimization algorithm is used to solve it, and the objective function is set as:

[0023]

[0024] Where M is the number of samples in the wind process, d i is the theoretical distance calculated according to the wind conditions and the point position, d is the distance between the two points actually measured, and the final delay time Δt is obtained by continuously adjusting the delay time Δt to minimize the objective function value;

[0025] The calculation formula of the acceleration factor is: Indicates the ratio of the wind speed change rate at point B relative to point A.

[0026] Preferably, the fluctuation of the power generation is analyzed, and the specific steps are as follows:

[0027] The calculation method of wind power fluctuation within the calculation time window Δt is defined as follows:

[0028]

[0029] The fluctuation of wind power within the time window Δt is calculated, and the frequency distribution histogram of power fluctuation is drawn. The normal distribution (ND) and kernel density estimation model (KDE) are used to fit it respectively. Since the power fluctuation is mostly concentrated near 0, the kernel density estimation model is used to calculate the lower confidence limit (LCL) and upper confidence limit (CI) of the power fluctuation at different confidence levels (CL), so as to characterize the detailed characteristics of the forward, reverse and overall power fluctuations. This indicator is suitable for different time and space scales.

[0030] Preferably, the complementarity of the generated power is analyzed, and the specific steps are as follows:

[0031] Sp1. Convert resource data of the area where the wind-solar combined power generation system is to be developed into power data, where wind power data is converted through the theoretical power curve of the corresponding wind turbine, and light power data is calculated through the formula.

[0032]

[0033] In the formula is the theoretical photovoltaic power at time t (MW), η is the conversion efficiency, A is the surface area of ​​the photovoltaic cell module (m 2 ), r(t) is the total solar radiation (W / m 2 ), T is the transmittance of the photovoltaic cell module, T(t) is the operating temperature of the photovoltaic cell module, which is replaced by the ambient temperature (℃);

[0034] Sp2. Normalize the theoretical wind power efficiency and photovoltaic power generation data obtained, and the calculation formula is:

[0035]

[0036] in is the normalized wind power at time t, is the wind power at time t, are the maximum and minimum values ​​of wind power time series respectively;

[0037] Sp3. Calculate the theoretical total power of the wind-solar combined power generation system at time t when the proportion of photovoltaic installed capacity is α and the proportion of wind power installed capacity is 1-α: P' h (t) = αP' s (t)+(1-α)P' w (t), where P' s (t) is the normalized theoretical photovoltaic power generation at time t, P' w (t) is the normalized theoretical wind power at time t,

[0038] Calculate the real-time complementarity rate of wind power and photovoltaic power in the joint power generation system from time t to time t+1

[0039]

[0040]

[0041]

[0042]

[0043] in, is the normalized fluctuation of the total theoretical wind and solar power power at time t; is the theoretical photovoltaic power fluctuation at time t after normalization, is the normalized theoretical wind power fluctuation at time t, and at the same time, the complementarity index (CICV) of theoretical wind power and photovoltaic power generation is calculated within the corresponding time analysis scale T.

[0044]

[0045] Where N is the number of samples; the complementary characteristics of wind and solar power generation under different installed capacity ratios are obtained through the above calculation.

[0046] Preferably, it also includes the step of sorting out the abnormal data types of the abnormal data in the collected wind and solar power station operation data, specifically: dividing the abnormal operation data of the wind and solar power station into four types, and classifying the abnormal data caused by measurement and communication equipment failures and existing in the form of continuous output during the actual operation of the wind and solar power station as type one; the abnormal data caused by measuring instrument errors and existing in the form of random dispersion during the actual operation of the wind and solar power station are classified as type two; the abnormal data caused by planned station maintenance, line failures and continuous energy abandonment and power rationing are classified as type three, among which type three has two forms of expression, one is accumulation in the form of horizontal lines, and the other form of expression is closely connected with normal data; the abnormal data caused by temporary station maintenance, deep participation of wind and solar power stations in power system dispatch and intermittent energy abandonment and power rationing factors are classified as type four.

[0047] Preferably, it also includes identifying and restoring abnormal data in the collected wind and solar power station operation data. The specific workflow is as follows:

[0048] The wind speed and irradiance are divided according to a certain value by using the longitudinal unilateral quartile method to obtain the power sequence in the meteorological interval, and the three quantiles of the power sequence in the meteorological interval are calculated. On this basis, the interquartile range is calculated, IQR = Q3-Q1 and the lower boundary L l =Q1-1.5×IQR. When the power in the meteorological interval is less than the lower boundary, it is regarded as abnormal data. The power is divided according to the percentage of installed capacity using the horizontal unilateral quartile method to obtain the meteorological sequence in the power interval. The three quantiles of the meteorological sequence in each power interval are calculated, and on this basis, the interquartile range and the upper boundary are calculated. L u =Q3+1.5×IQR, when the wind speed and irradiance within the power range are greater than the upper boundary, they are regarded as abnormal data.

[0049] Preferably, the method of correcting the wind speed / irradiance error of numerical weather forecast using time series transmission specifically comprises the following steps:

[0050] Sp1, NWP wind speed correction, the input variable set consists of the NWP wind speed at time t+1 and the measured wind speed sequence at time t, and the output variable set consists of the measured wind speed sequence at time t+1; NWP irradiance correction, the input variable set consists of the NWP irradiance at time t+1, the time label at time t+1 and the measured irradiance sequence at time t, and the output variable set consists of the measured irradiance sequence at time t+1;

[0051] Sp2, 1m / s and 100W / m 2 The NWP wind speed and irradiance data are segmented for intervals, and then the input variable set and the output variable set are simultaneously divided into k interval segments;

[0052] Sp3. Establish a mapping relationship between input variables and output variables in each interval, train wind speed and irradiance correction models, and store model parameters in each interval for future use in correcting NWP wind speed / irradiance;

[0053] Sp4, based on the NWP wind speed / irradiance at time t+1, the model required for NWP correction is determined in real time, and the rolling iteration mode is used for correction. Within the correction time n, except for the first correction point, the correction value of the previous moment is substituted into the model input end as the measured value for NWP correction, and the NWP correction results under each correction time are obtained. Based on this, a database of NWP wind speed / irradiance correction results under different time lengths is established;

[0054] Sp5. Based on the above-obtained correction result database, construct the power prediction input and output variable sets. For wind power prediction, the input variable set consists of the NWP corrected wind speed at time t+1 and the NWP wind direction sequence at time t+1, and the output variable set consists of the measured wind power sequence at time t+1; for photovoltaic power prediction, the input variable set consists of the NWP corrected irradiance at time t+1, the NWP temperature at time t+1, and the time label sequence at time t+1, and the output variable set consists of the measured photovoltaic power sequence at time t+1;

[0055] Sp6. Divide each input variable set and output variable set by month based on time, establish a wind power / photovoltaic power generation prediction model in each month, and obtain wind power / photovoltaic power generation prediction results at different time scales.

[0056] Preferably, the wind and solar power generation power and load joint prediction based on deep learning includes the following steps:

[0057] Adopt variable attention mechanism to extract key input information of wind and solar load;

[0058] Based on multi-task learning and fully connected neural network, wind power prediction tasks and optical power prediction tasks are carried out simultaneously, and homoscedastic uncertainty is introduced to achieve automatic optimization of loss weights for different prediction tasks;

[0059] Based on the long short-term memory neural network, a load forecasting model considering future wind and solar power generation information is established to predict the power of each wind farm and the total power of regional wind farms, the power of each photovoltaic power station and the total power of regional photovoltaic power stations.

[0060] Preferably, the specific steps of the wind-solar power generation and load joint prediction based on deep learning are as follows:

[0061] Construct the input variable set and output variable set required for wind power prediction and photovoltaic power prediction. The input variable set consists of the numerical weather forecast of each wind farm and each photovoltaic power station, including wind speed, irradiance and temperature. The output variable set consists of the power of each wind farm and each photovoltaic power station, and the total power of regional wind farms and regional photovoltaic power stations.

[0062] Multi-task learning is implemented using hard parameter sharing. The hidden layer is shared among all tasks, and the output layer of each task is constructed separately. The total loss function of multi-task learning is

[0063]

[0064] Where L total is the total loss function of each task, w i is the weight of the i-th task, L i is the loss function of the i-th task. The Bayesian modeling idea is used to automatically obtain the optimal weight of each task. The loss functions of wind power prediction task and photovoltaic power prediction task are constructed based on homoscedastic uncertainty.

[0065] A fully connected neural network is used to build a shared hidden layer in the joint prediction of wind power and photovoltaic power generation. The quantile regression algorithm is used to realize the uncertainty prediction of wind and solar power generation. The constructed wind-solar joint prediction model is used to predict the power of each wind farm and the total power of regional wind farms, the power of each photovoltaic power station and the total power of regional photovoltaic power stations.

[0066] The present invention has the following beneficial effects:

[0067] In the present invention, the system significantly improves the quality of data and the accuracy of analysis through a series of sophisticated data processing and analysis methods. In the data collection link, the reasonable collection frequency setting and the comprehensive collection of various data provide a rich source of information for subsequent analysis. In view of the problem that data is easily affected in cold environments, the preprocessing methods such as the triple standard deviation method to eliminate outliers and the linear interpolation method to fill missing values ​​are adopted. Although there may be certain limitations in complex situations, they effectively clean up the noise in the data in conventional scenarios, making the basic data more reliable. The establishment of the abnormal data type combing unit is a highlight. The abnormal data is classified in detail and a targeted identification process is formulated. Whether it is equipment failure, instrument error or various abnormal situations related to the operation plan, it can be accurately judged. This not only improves the accuracy of the data, but also provides a clear direction for subsequent targeted equipment maintenance, calibration and operation strategy adjustment. In the analysis of power generation characteristics, the time shift analysis accurately grasps the time relationship of wind speed through precise algorithms and consideration of practical factors; the multi-model application and indicator system of volatility analysis comprehensively characterize the power fluctuation characteristics; and the complementarity analysis provides strong support for the planning of joint power generation systems.

[0068] The system in the present invention performs well in improving prediction accuracy and overall performance. The application of multi-task learning and variable attention mechanism makes the joint prediction of wind and photovoltaic power generation and load more accurate and efficient. By considering the relationship between multiple factors and tasks at the same time, the model can better adapt to the complex characteristics of wind and photovoltaic power generation in cold regions, extract key information, and optimize prediction results. The innovative method of the time series transfer correction module introduces the time series transfer relationship and performs multi-step correction and modeling, which greatly improves the mapping accuracy between input and output variables and brings higher accuracy to wind power and photovoltaic power generation prediction.

[0069] The establishment of the resource database and the resource allocation plan based on it in the present invention is one of the important advantages of the system. The detailed information of human resources, material resources and financial resources can be mastered in real time, so that the system can respond quickly when facing various emergencies. Whether it is equipment damage caused by natural disasters or power supply shortage, the corresponding resources can be allocated in a timely and accurate manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 Establishing an overall system configuration diagram for the adaptive prediction model of the present invention;

[0071] Figure 2 The data cleaning module workflow diagram of the present invention;

[0072] Figure 3 It is a schematic diagram of the overall process of the present invention. DETAILED DESCRIPTION

[0073] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] Example 1

[0075] like Figure 1-3 As shown, a multi-scale fusion prediction system for cold-region wind and solar power generation prediction includes a power generation characteristic analysis module, a data cleaning module, a timing transfer correction module, and a wind and solar power generation and load joint prediction module. The power generation characteristic analysis module performs time shift analysis, power generation volatility analysis, and complementarity analysis of cold-region wind and solar resources and power generation in sequence. The data cleaning module includes an abnormal data type combing unit and an abnormal data identification and restoration unit; the timing transfer correction module adopts NWP wind speed and irradiance timing transfer correction, and the wind and solar power generation and load joint prediction module adopts multi-task learning for wind and solar power generation and load joint prediction.

[0076] A multi-scale fusion prediction method for wind and solar power generation in cold regions comprises the following steps:

[0077] Based on the measured wind and solar input and power output data of wind and solar power generation base generators in different regions and time scales under different wind speeds, wind directions, and light intensities, the characteristics of the time shift, volatility, and complementarity of the power generation of wind and solar power generation base generators are analyzed to build a power characteristics database;

[0078] Use time series transfer to correct the wind speed / irradiance error of numerical weather forecasts and establish a wind and solar power generation power prediction model;

[0079] Based on deep learning, the wind and solar power generation power and load joint prediction is carried out, and the wind and solar power generation prediction models in different regions are regionally combined for multi-scale fusion prediction.

[0080] Example 2

[0081] like Figure 1-3 As shown in the figure, when analyzing the time shift of wind and solar resources and power generation in cold regions, the specific steps are as follows:

[0082] Sp1. When analyzing the time shift, first collect the wind speed sequences of wind turbines at different locations in the same wind farm. When collecting the sequences, collect them through anemometers and meteorological towers. After collecting the wind speed sequences of wind turbines at different points, preprocess them. During preprocessing, first perform a quality check on the original wind speed data to remove outliers and missing values. The three-times standard deviation method is used to determine the outliers. When determining, if the difference between a single data point and the mean exceeds three times the standard deviation, the data point is considered to be an outlier. For missing values, linear interpolation is used to fill them. Linear estimation is performed based on the values ​​of adjacent data points to supplement the missing values. When linearly estimating, let the wind speed sequence be V=[v1,v2,...,vn], let vi be a missing value, then its adjacent data point is v i-1 and v i+1 In the actual process, when the missing value is not the first and last data point of the sequence but a boundary point, only one adjacent point is taken for special processing, and the linear estimation formula is: By linear estimation formula v i Positioned at v i-1 and v i+1 The midpoint position between the ranges, when the time interval is considered in the calculation process, let the time interval be Δt, let the adjacent time points t in the time series be i-1 ,t i ,t i+1 , the corresponding wind speed values ​​are v i-1 、v i (missing values) and v i+1 ,but

[0083] In a wind speed time series, assuming the time interval is 10 minutes, we know that at t1 = 10:00, the wind speed v1 = 5m / s, at t3 = 10:20, the wind speed v3 = 7m / s, and at t2 = 10:10, the wind speed value is missing. According to the above formula, we can get:

[0084]

[0085] Sp2. Divide the wind speed and wind direction sequences according to fixed time windows to obtain a set of wind processes. By judging the relative position of the main wind direction and the point in the wind process, select a new wind speed sequence that can be studied for time shift. Define the delay time (Δt) and acceleration factor (Δv) indicators. Take the actual distance between the two points and the actual wind conditions as constraints, and minimize the Euclidean distance between the corresponding wind processes as the optimization goal. Use the firefly optimization algorithm to solve, where the delay time represents the time lag of the wind speed change at one point relative to another point, and the acceleration factor reflects the difference in the rate of change of wind speed at different points. The specific calculation method is as follows:

[0086] Assume that there are two points A and B, whose wind speed sequences are vA (t) and v B (t), t represents time. First, the delay time is preliminarily estimated by the cross-correlation function, and v is calculated. A (t) and v B The cross-correlation function of (t):

[0087]

[0088] Where τ is the time delay, N is the number of data points, and the τ value corresponding to the maximum value of the cross-correlation function is the initial estimated delay time Δt0.

[0089] Then, the initially estimated delay time is optimized, considering the actual distance d between the two points and the actual wind conditions (average wind speed and wind direction changes) as constraints, and the minimum Euclidean distance between the corresponding wind processes is taken as the optimization goal. The firefly optimization algorithm is used to solve it, and the objective function is set as:

[0090]

[0091] Where M is the number of samples in the wind process, d i is the theoretical distance calculated according to the wind conditions and point positions, d is the actual measured distance between the two points, and the final delay time Δt is obtained by continuously adjusting the delay time Δt to minimize the objective function value.

[0092] The calculation formula of the acceleration factor is: Indicates the ratio of the wind speed change rate at point B relative to point A.

[0093] Sp3. Based on the delay time and acceleration factor of each wind process, the new wind speed sequence after time shift is considered, and the correlation between the wind speed sequences at different spatial points before and after the time shift is compared and analyzed. The indicators for evaluating the time shift are selected according to the distance and wind power installed capacity. When the distance between wind farms exceeds 60km, it is determined that there is no acceleration factor influence between the wind speed sequences at the locations of each wind farm.

[0094] Example 3

[0095] like Figure 1-3 As shown in the figure, when analyzing the volatility of wind and solar resources and power generation in cold regions, the specific steps are as follows:

[0096] The calculation method of wind power fluctuation within the calculation time window Δt is defined as follows:

[0097]

[0098] The fluctuation of wind power within the time window Δt is calculated, and the frequency distribution histogram of power fluctuation is drawn. The normal distribution (ND) and kernel density estimation model (KDE) are used to fit it respectively. Since the power fluctuation is mostly concentrated near 0, the kernel density estimation model is used to calculate the lower confidence limit (LCL) and upper confidence limit (CI) of the power fluctuation at different confidence levels (CL) to characterize the detailed characteristics of the forward, reverse and overall power fluctuations. This indicator is suitable for different time and space scales, and the calculation time window and confidence level can be selected according to needs.

[0099] Example 4

[0100] like Figure 1-3 As shown in the figure, when analyzing the complementarity between wind and solar resources and power generation in cold regions, the specific steps are as follows:

[0101] Sp1. Convert resource data of the area where the wind-solar combined power generation system is to be developed into power data, where wind power data is converted through the theoretical power curve of the corresponding wind turbine, and light power data is calculated through the formula.

[0102]

[0103] In the formula is the theoretical photovoltaic power at time t (MW), η is the conversion efficiency, A is the surface area of ​​the photovoltaic cell module (m 2 ), r(t) is the total solar radiation (W / m 2 ), T is the transmittance of the photovoltaic cell module, T(t) is the operating temperature of the photovoltaic cell module, which is replaced by the ambient temperature (℃);

[0104] Sp2. Normalize the theoretical wind power efficiency and photovoltaic power generation data obtained, and the calculation formula is:

[0105]

[0106] in is the normalized wind power at time t, is the wind power at time t, are the maximum and minimum values ​​of wind power time series respectively;

[0107] Sp3. Calculate the theoretical total power of the wind-solar combined power generation system at time t when the proportion of photovoltaic installed capacity is α and the proportion of wind power installed capacity is 1-α: P' h (t) = αP's (t)+(1-α)P' w (t), where P' s (t) is the normalized theoretical photovoltaic power generation at time t, P' W (t) is the normalized theoretical wind power at time t,

[0108] Calculate the real-time complementarity rate of wind power and photovoltaic power in the joint power generation system from time t to time t+1

[0109]

[0110]

[0111]

[0112]

[0113] in, is the normalized fluctuation of the total theoretical wind and solar power power at time t; is the theoretical photovoltaic power fluctuation at time t after normalization, is the theoretical power fluctuation of wind power at time t after normalization, and at the same time, the complementarity index (Complementarity Index Considering Volatility, CICV) of theoretical wind power and photovoltaic power generation is calculated within the corresponding time analysis scale T.

[0114]

[0115] Where N is the number of samples. Through the above calculation, the complementary characteristics of wind and solar power generation under different installed capacity ratios are obtained, and then it is found that at what wind and solar installed capacity ratio the complementarity of the combined power generation system is the strongest, providing effective assistance for the planning of the combined power generation system.

[0116] Example 5

[0117] like Figure 1-3As shown, it also includes a data cleaning module, which includes an abnormal data type combing unit and an abnormal data identification and restoration unit, wherein the specific work flow of the abnormal data type combing unit includes: dividing the abnormal operation data of the wind and solar power station into four types: the abnormal data caused by measurement and communication equipment failures and existing in the form of continuous output during the actual operation of the wind and solar power station are classified as type one (accumulated abnormal data above the power curve); the abnormal data caused by errors in measuring instruments (such as wind speed sensors, irradiance sensors) and existing in a randomly dispersed form during the actual operation of the wind and solar power station are classified as type two (dispersed abnormal data above the power curve); the abnormal data caused by planned station maintenance, line failures and continuous power abandonment and power restriction are classified as type three (accumulated abnormal data below the power curve), wherein type three has two forms of expression, one is accumulation in the form of horizontal lines, and the other form is closely connected with normal data; the abnormal data caused by temporary station maintenance, deep participation of wind and solar power stations in power system scheduling, and short-term intermittent power abandonment and power restriction factors are classified as type four (dispersed abnormal data below the power curve).

[0118] The specific sorting process is as follows:

[0119] The goal of sorting out abnormal data types is to clarify the different types and characteristics of abnormal operation data of wind and solar power stations, providing a basis for the accurate identification and effective processing of subsequent abnormal data. Through the detailed classification of abnormal data types, corresponding processing strategies can be adopted more targetedly to improve the accuracy and reliability of data, thereby ensuring the stable operation of wind and solar power stations and the effectiveness of data analysis.

[0120] Specific workflow planning,

[0121] Type 1: Abnormal data accumulated above the power curve (equipment failure type).

[0122] 1. Data monitoring and collection: Establish a real-time data monitoring system to continuously collect the operating data of wind and solar power stations, focus on power data and related measurement data (such as wind speed, irradiance) and communication status data, and set the data collection frequency, such as collecting data every 5 minutes or every 10 minutes, to ensure that possible abnormal situations can be captured in time.

[0123] 2. Identify the characteristics of abnormal data and formulate the criteria for determining the accumulation type abnormal data above the power curve. When the power values ​​of multiple consecutive data points are significantly higher than the expected range of the normal power curve and show an accumulation trend, it is preliminarily determined to be possible type 1 abnormal data. For example, under normal circumstances, the power of a wind turbine at a certain wind speed should fluctuate within a reasonable range. If the power value exceeds the upper limit of the range by more than 50% for multiple consecutive moments, and these abnormal data are adjacent in time series and are large in number (such as more than 10 consecutive data points), it meets the characteristics of the accumulation type above the power curve.

[0124] 3. Troubleshooting and classification of fault causes. Once suspected type 1 abnormal data is found, immediately start the fault cause troubleshooting process. First, check the working status of the measuring equipment (wind speed sensor, irradiance sensor, etc.), check whether the equipment has error information, whether the sensor is working properly (such as whether it is blocked, damaged, etc.), and check the communication equipment at the same time to check whether the data transmission link is unobstructed and whether there are abnormal conditions such as packet loss and delay. If it is determined that the data abnormality is caused by a failure of the measuring equipment or communication equipment, it will be clearly classified as type 1, and the relevant information of the faulty equipment will be recorded, including the equipment name, number, failure time, etc., for subsequent maintenance and data analysis.

[0125] Type 2: Scattered abnormal data above the power curve (instrument error type);

[0126] Data statistics and analysis: regularly conduct statistical analysis on the operating data of wind and solar power stations, calculate statistical indicators such as the mean and standard deviation of power data, and establish a normal distribution model for power data. For example, perform statistical analysis on power data within a week to obtain the mean value μ and standard deviation σ of power.

[0127] Abnormal data detection: During real-time data monitoring, when the power value of a single or a few data points deviates from the normal power curve and is located above the power curve, and at the same time meets the condition of exceeding the average value μ plus a certain multiple of the standard deviation σ (such as 2σ, 3σ), it is listed as suspected type II abnormal data. For example, at a certain moment, the wind turbine power value is P, if P>μ+2σ, and the data point is relatively isolated in the time series, and the surrounding data points are all within the normal range, it may be type II abnormal data.

[0128] Instrument error assessment and classification: conduct error assessment on the measuring instruments corresponding to the suspected abnormal data. Check the calibration records of the instrument to see if the calibration cycle has been exceeded. Compare the measurement data of different instruments at the same measurement point (if there are multiple sensors of the same type) to determine whether there are abnormal data in individual instruments. If it is determined that the data abnormality is caused by the error of the measuring instrument itself, classify it as type II. Record relevant information of the instrument, including instrument type, calibration status, etc., in order to arrange calibration or maintenance of the instrument.

[0129] Type 3: Accumulated abnormal data below the power curve (planned maintenance and power restriction type);

[0130] Operation plan and data association: Establish an operation plan database for wind and solar power stations, including planned station maintenance plans, line fault maintenance plans, and energy curtailment and power restriction plans. Associate the operation plan data with the real-time operation data to promptly identify abnormal data related to the plan. For example, record the start time, end time, and maintenance equipment type of the station maintenance in the maintenance plan. When abnormal data with accumulation below the power curve appears in the real-time data and the time is consistent with the maintenance plan time, further analysis is performed.

[0131] Identification and classification of abnormal data features. For the accumulation type abnormal data below the power curve, they are classified according to their manifestation. If the abnormal data is accumulated in a relatively regular horizontal line form, and the duration matches the planned maintenance time or power restriction time, it is classified as type three horizontal line accumulation subtype. For example, if the power data is continuously stable at a low level during a certain period, which is approximately a horizontal line, and the period is within the planned station maintenance time period, it is identified as this subtype. If the abnormal data accumulates below the power curve, but there is a certain transition with the normal data, and this situation occurs during the planned maintenance, line failure or continuous power curtailment, it is classified as type three accumulation subtype connected to normal data. For example, during the power curtailment period, the power data gradually decreases from the normal level and accumulates at a lower level, and gradually returns to normal after the power restriction ends, and has a clear transition trend with the normal data, it belongs to this subtype. Record the relevant information of each subtype of abnormal data, such as the start time, duration and accumulation degree.

[0132] Type 4: Dispersed abnormal data below the power curve (temporary maintenance and scheduling type);

[0133] Event monitoring and data marking: Establish a wind and solar power plant event monitoring system to monitor temporary maintenance events at the site, wind and solar power plants participating in power system dispatch events, and short-term power curtailment and power restriction events in real time. When these events occur, mark the operating data of the corresponding time period. For example, when receiving a notice of temporary maintenance at the site, immediately mark the data from the start to the end of the maintenance in the data system for subsequent analysis.

[0134] Abnormal data identification and classification In the marked data, find the scattered abnormal data below the power curve. If during the event, the power value of a single or a few data points is lower than the normal power curve, and the distribution is relatively scattered, it is identified as suspected type IV abnormal data. For example, during the period when the wind and solar power stations are deeply involved in the power system dispatch, the power data at certain moments suddenly drops and is lower than the normal range, and is scattered in the time series, which is significantly different from the surrounding normal data points. It may be type IV abnormal data. Through the correlation analysis with the event records, it is determined that these abnormal data are caused by temporary maintenance of the station, wind and solar power stations participating in the power system dispatch, or short-term power abandonment and power restriction factors, and they are classified as type IV. Record the detailed information of the abnormal data, including the correlation relationship with the event, the specific time and power value of the data point, etc. Through the above detailed planned workflow, the abnormal data type combing unit can accurately classify the abnormal operation data of the wind and solar power station, and provide strong support for the subsequent abnormal data identification and restoration unit, thereby improving the effect and quality of data cleaning, and ensuring the reliability and availability of wind and solar power station operation data.

[0135] Example 6

[0136] like Figure 1-3 As shown, the specific workflow of the abnormal data identification and restoration unit is as follows:

[0137] The wind speed and irradiance are set to 0.5 m / s and 20 W / m by using the longitudinal unilateral quartile method. 2 Divide it to get the power sequence in the meteorological interval, calculate the three quantiles of the power sequence in the meteorological interval, and on this basis calculate the interquartile range, IQR = Q3-Q1 and the lower boundary L l =Q1-1.5×IQR. When the power in the meteorological interval is less than the lower boundary, it is regarded as abnormal data. The power is divided by 2% of the installed capacity using the horizontal unilateral quartile method to obtain the meteorological sequence in the power interval. The three quantiles of the meteorological sequence in each power interval are calculated, and on this basis, the interquartile range and the upper boundary are calculated. L u =Q3+1.5×IQR, when the wind speed and irradiance within the power range are greater than the upper boundary, they are regarded as abnormal data.

[0138] Example 7

[0139] like Figure 1-3 As shown in the figure, the timing transfer correction module adopts the timing transfer correction method of NWP wind speed and irradiance. In addition to the NWP data at time t+1, the input variables of this method also include the measured data at time t. By introducing the transfer relationship in the wind speed and irradiance time series, the certainty of the mapping relationship between the input variables and the output variables is improved. It is applied to the power prediction of wind farms and photovoltaic power stations, and specifically includes the following steps:

[0140] Sp1. First, when the NWP wind speed is corrected, the input variable set consists of the NWP wind speed at time t+1 and the measured wind speed sequence at time t, and the output variable set consists of the measured wind speed sequence at time t+1; when the NWP irradiance is corrected, the input variable set consists of the NWP irradiance at time t+1, the time label at time t+1 and the measured irradiance sequence at time t, and the output variable set consists of the measured irradiance sequence at time t+1;

[0141] Sp2, then segment the data to 1m / s and 100W / m 2 The NWP wind speed and irradiance are segmented for intervals, and then the input variable set and the output variable set are simultaneously divided into k segments;

[0142] Sp3, then establish mapping relationship and model training. In each interval, the mapping relationship between input variables and output variables is established, the wind speed and irradiance correction model is trained, and the model parameters in each interval are stored for future use in correcting NWP wind speed / irradiance;

[0143] Sp4, then carry out rolling iterative correction, based on the NWP wind speed / irradiance at time t+1, determine the model required for NWP correction in real time, and use rolling iterative mode for correction. Within the correction time n, except for the first correction point, the correction value of the previous moment is substituted into the model input end as the measured value to perform NWP correction during correction, and the NWP correction results under each correction time are obtained, and a database of NWP wind speed / irradiance correction results under different time lengths is established accordingly;

[0144] Sp5. Construct power prediction input and output variable sets. For wind power prediction, the input variable set consists of the NWP corrected wind speed at time t+1 and the NWP wind direction at time t+1, and the output variable set consists of the measured wind power sequence at time t+1. For photovoltaic power prediction, the input variable set consists of the NWP corrected irradiance at time t+1, the NWP temperature at time t+1, and the time label sequence at time t+1, and the output variable set consists of the measured photovoltaic power sequence at time t+1.

[0145] Sp6. Finally, each input variable set and output variable set are divided into months based on time, and a wind power / photovoltaic power generation prediction model is established in each month to obtain wind power / photovoltaic power generation prediction results at different time scales.

[0146] Example 8

[0147] like Figure 1-3As shown in the figure, the joint prediction of wind and solar power generation power and load based on deep learning includes the following steps: using the variable attention mechanism to extract the key input information of wind, solar and load; constructing the input variable set and output variable set required for wind power prediction and photovoltaic power generation power prediction. The input variable set consists of the numerical weather forecast of each wind farm and each photovoltaic power station, including wind speed, irradiance and temperature. The output variable set consists of the power of each wind farm and each photovoltaic power station, and the total power of regional wind farms and regional photovoltaic power stations.

[0148] Based on multi-task learning and fully connected neural network, wind power prediction tasks and optical power prediction tasks are carried out simultaneously, and homoscedastic uncertainty is introduced to realize automatic optimization of loss weights of different prediction tasks; multi-task learning is realized by hard parameter sharing, hidden layers are shared among all tasks, and the output layer of each task is constructed separately. The total loss function of multi-task learning is:

[0149]

[0150] Where L total is the total loss function of each task, w i is the weight of the i-th task, L i is the loss function of the i-th task. The Bayesian modeling idea is used to automatically obtain the optimal weight of each task, and the loss functions of wind power prediction task and photovoltaic power prediction task are constructed based on homoscedastic uncertainty.

[0151] The loss function construction process of the wind power prediction task and photovoltaic power prediction task based on homoscedastic uncertainty is as follows. First, the likelihood function of the model is defined as shown in the formula. For multiple tasks, its likelihood function is shown in the formula:

[0152] p(y|f W (x)) = N(f W (x),σ 2 )

[0153]

[0154] Where x is the model input, f w (x) is the model output, W is the task weight parameter, σ is the model observation noise parameter; y1,...,y k Output for model task 1,…,task k.

[0155] The maximum likelihood estimation is converted into minimizing the negative log-likelihood. The outputs of the wind power prediction task and the photovoltaic power prediction task are represented by y1 and y2 respectively. The logarithm of the likelihood function of the two tasks is taken, and the loss function of the optimization model is converted into minimizing the negative log-likelihood, as shown in the formula:

[0156]

[0157] Where L1(W) is the predicted loss of wind power, and L2(W) is the predicted loss of photovoltaic power.

[0158] Based on the long short-term memory neural network, a load forecasting model considering future wind and solar power generation information is established to predict the power of each wind farm and the total power of regional wind farms, the power of each photovoltaic power station and the total power of regional photovoltaic power stations; that is, a fully connected neural network is used to build a shared hidden layer in the joint prediction of wind power and photovoltaic power generation, and the quantile regression algorithm is used to realize the uncertainty prediction of wind and solar power generation. The constructed wind-solar joint prediction model is used to predict the power of each wind farm and the total power of regional wind farms, the power of each photovoltaic power station and the total power of regional photovoltaic power stations. Specific embodiment 9

[0160] like Figure 1-3 As shown, the system encrypts the collected data and predicted data, and uses the AES symmetric encryption algorithm and the RSA asymmetric encryption algorithm to encrypt the data in a combined manner. During the data collection stage, the AES symmetric encryption algorithm is used to quickly encrypt the original cold-region wind and solar power generation data.

[0161] In the data collection stage, the original cold-region wind and solar power generation data is quickly encrypted using the AES symmetric encryption algorithm due to its large data volume. Specifically, the system generates an independent AES symmetric encryption key for each data collection point, and uses the key to encrypt the collected data in real time. For example, the data collection device collects wind and solar power generation data at a certain interval (such as every second), encrypts the data using the AES key at the moment of collection, and then transmits the encrypted data to the system's data storage center.

[0162] The system establishes a resource database for the human resources (technicians and maintenance workers), material resources (spare equipment, maintenance tools and emergency supplies) and financial resources within the cold-region wind and solar power generation system, and keeps track of the quantity, location and status of resources in real time. It formulates a resource allocation plan based on the type and scale of the emergency. When equipment is damaged by natural disasters, it promptly deploys maintenance personnel and spare equipment to the disaster site for emergency repairs. When the power supply is tight, it deploys financial resources to purchase emergency power generation equipment and pay for additional power transmission costs. It establishes a coordination mechanism for resource allocation and establishes close ties with suppliers and partners.

[0163] The specific resource allocation plan and calculation formula of the above coordination and allocation mechanism are as follows:

[0164] (1) Calculation of human resource deployment in the event of equipment failure

[0165] Fault urgency evaluation index (F):

[0166] Calculation formula:

[0167] Where T downtime is the time the device has been down, T max_downtime It is the maximum tolerable downtime of the equipment set according to the importance of the equipment and the power supply demand. loss It is the power loss rate caused by equipment shutdown, which can be calculated by dividing the difference between the current power generation and the normal power generation by the normal power generation, that is,

[0168] For example, if the normal power generation of a wind turbine is 1000kw and the current power generation is 600kw, then

[0169]

[0170] If the maximum tolerable downtime of the fan is 2 hours and it has been down for 1 hour, then the fault urgency is

[0171] (2) Maintenance personnel suitability index (M)

[0172] Calculation formula:

[0173]

[0174] Where n is the number of skills that maintenance personnel possess, s i is the maintenance personnel's proficiency rating for the i-th skill (the value range is 0-1, 1 means very proficient), E i It is the importance weight of the skill required for equipment failure (determined according to the failure type and maintenance history data, the sum is 1).

[0175] For example, a maintenance personnel has three skills for fan maintenance, namely electrical maintenance skills (proficiency score 0.8, importance weight 0.4), mechanical maintenance skills (proficiency score 0.7, importance weight 0.3) and control system maintenance skills (proficiency score 0.6, importance weight 0.3). The fitness of the maintenance personnel is M = 0.8 × 0.4 + 0.7 × 0.3 + 0.6 × 0.3 = 0.71.

[0176] (3) Human resource deployment decisions

[0177] Select the maintenance personnel with the highest fitness and who is currently idle or available to go to the fault site. When there are multiple maintenance personnel with similar fitness, the maintenance personnel closest to the faulty equipment will be selected first to reduce travel time. The distance calculation can be based on the Euclidean distance formula in the Geographic Information System (GIS):

[0178]

[0179] Where (x1, x1) is the geographic location coordinates of the faulty equipment, and (x2, x2) is the geographic location coordinates of the maintenance personnel.

[0180] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprising a reference structure" do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0181] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-scale fusion prediction method for wind and solar power generation in cold regions, characterized in that: The steps include: Based on the measured wind and solar input and power output data of wind and solar power generation base generators in different regions and time scales under different wind speeds, wind directions, and light intensities, the characteristics of the time shift, volatility, and complementarity of the power generation of wind and solar power generation base generators are analyzed to build a power characteristics database; Use time series transfer to correct the wind speed / irradiance error of numerical weather forecasts and establish a wind and solar power generation power prediction model; Based on deep learning, the wind and solar power generation power and load joint prediction is carried out, and the wind and solar power generation prediction models in different regions are regionally combined for multi-scale fusion prediction.

2. The multi-scale fusion prediction method for wind and solar power generation in cold regions according to claim 1 is characterized in that: The specific steps of analyzing the time shift of the generated power are as follows: Sp1. When performing the time-shift analysis, the wind speed sequences of wind turbines at different locations in the same wind farm are first collected. The sequences are collected through anemometers and meteorological towers. After the wind speed sequences of wind turbines at different locations are collected, they are preprocessed. During the preprocessing, the quality of the original wind speed data is first checked to remove abnormal values ​​and missing values. The abnormal values ​​are determined by the triple standard deviation method. When determining, if the difference between a single data point and the mean exceeds three times the standard deviation, the data point is considered to be an abnormal value. For missing values, linear interpolation is used to fill them. Linear estimation is performed based on the values ​​of adjacent data points to supplement the missing values. When linearly estimating, the wind speed sequence is set to V = [v1, v2, ..., vn], and vi is set to be a missing value, then its adjacent data point is v i-1 and v i+1 In the actual process, when the missing value is not the first and last data point of the sequence but a boundary point, only one adjacent point is taken for special processing, and the linear estimation formula is: By linear estimation formula v i Positioned at v i-1 and v i+1 The midpoint position between the ranges, when the time interval is considered in the calculation process, let the time interval be Δt, let the adjacent time points t in the time series be i-1 ,t i ,t i+1 , the corresponding wind speed values ​​are v i-1 、v i (missing values) and v i+1 ,but Sp2. Divide the wind speed and wind direction sequences according to fixed time windows to obtain a set of wind processes. By judging the relative position of the main wind direction and the point in the wind process, select a new wind speed sequence for studying time shift, define the delay time (Δt) and acceleration factor (Δv) indicators, where the delay time indicates the time lag of the wind speed change at one point relative to another point, and the acceleration factor reflects the difference in the rate of change of the wind speed at different points. The actual distance between the two points and the actual wind conditions are constrained, and the minimum Euclidean distance between the corresponding wind processes is the optimization goal. The firefly optimization algorithm is used to solve the above indicators; Sp3. Based on the delay time and acceleration factor of each wind process obtained, the time shift between wind speed sequences at different spatial points is analyzed, and at the same time, indicators for evaluating time shift are selected according to distance and wind power installed capacity. When the distance between wind farms exceeds 60km, it is determined that there is no acceleration factor influence between the wind speed sequences at the locations of each wind farm.

3. The multi-scale fusion prediction method for wind and solar power generation in cold regions according to claim 2 is characterized in that: The firefly optimization algorithm is as follows: Assume that there are two points A and B, whose wind speed sequences are v A (t) and v B (t), t represents time. First, the delay time is preliminarily estimated by the cross-correlation function, and v is calculated. A (t) and v B The cross-correlation function of (t): Where τ is the time delay, N is the number of data points, and the τ value corresponding to the maximum value of the cross-correlation function is the initial estimated delay time Δt0; Then, the initially estimated delay time is optimized, considering the actual distance d between the two points and the actual wind conditions of the average wind speed and wind direction changes as constraints, and taking the minimum Euclidean distance between the corresponding wind processes as the optimization goal. The firefly optimization algorithm is used to solve it, and the objective function is set as: Where M is the number of samples in the wind process, d i is the theoretical distance calculated according to the wind conditions and the point position, d is the distance between the two points actually measured, and the final delay time Δt is obtained by continuously adjusting the delay time Δt to minimize the objective function value; The calculation formula of the acceleration factor is: Indicates the ratio of the wind speed change rate at point B to that at point A.

4. The multi-scale fusion prediction method for wind and solar power generation in cold regions according to claim 1 is characterized in that: The specific steps of analyzing the fluctuation of power generation are as follows: The calculation method of wind power fluctuation within the calculation time window Δt is defined as follows: The fluctuation of wind power within the time window Δt is calculated, and the frequency distribution histogram of power fluctuation is drawn. The normal distribution (ND) and kernel density estimation model (KDE) are used to fit it respectively. Since the power fluctuation is mostly concentrated near 0, the kernel density estimation model is used to calculate the lower confidence limit (LCL) and upper confidence limit (CI) of the power fluctuation at different confidence levels (CL), so as to characterize the detailed characteristics of the forward, reverse and overall power fluctuations. This indicator is suitable for different time and space scales.

5. The multi-scale fusion prediction method for wind and solar power generation in cold regions according to claim 1 is characterized in that: The specific steps of analyzing the complementarity of power generation are as follows: Sp1. Convert resource data of the area where the wind-solar combined power generation system is to be developed into power data, where wind power data is converted through the theoretical power curve of the corresponding wind turbine, and light power data is calculated through the formula. In the formula is the theoretical photovoltaic power at time t (MW), η is the conversion efficiency, A is the surface area of ​​the photovoltaic cell module (m 2 ), r(t) is the total solar radiation (W / m 2 ), T is the transmittance of the photovoltaic cell module, T(t) is the operating temperature of the photovoltaic cell module, which is replaced by the ambient temperature (℃); Sp2. Normalize the theoretical wind power efficiency and photovoltaic power generation data obtained, and the calculation formula is: in is the normalized wind power at time t, is the wind power at time t, are the maximum and minimum values ​​of wind power time series respectively; Sp3. Calculate the theoretical total power of the wind-solar combined power generation system at time t when the proportion of photovoltaic installed capacity is α and the proportion of wind power installed capacity is 1-α: P' hh (t) = αP' s (t)+(1-α)P' w (t), where P' s (t) is the normalized theoretical photovoltaic power generation at time t, P' w (t) is the normalized theoretical wind power at time t, Calculate the real-time complementarity rate of wind power and photovoltaic power in the joint power generation system from time t to time t+1 in, is the normalized fluctuation of the total theoretical wind and solar power power at time t; is the theoretical photovoltaic power fluctuation at time t after normalization, is the normalized theoretical wind power fluctuation at time t, and at the same time, the complementarity index (CICV) of theoretical wind power and photovoltaic power generation is calculated within the corresponding time analysis scale T. Where N is the number of samples; the complementary characteristics of wind and solar power generation under different installed capacity ratios are obtained through the above calculation.

6. The multi-scale fusion prediction method for wind and solar power generation in cold regions according to claim 1 is characterized in that: It also includes the step of sorting out the abnormal data types of the abnormal data in the collected wind and solar power station operation data, specifically: dividing the abnormal operation data of the wind and solar power station into four types, and classifying the abnormal data caused by measurement and communication equipment failures and existing in the form of continuous output during the actual operation of the wind and solar power station as type one; the abnormal data caused by measurement instrument errors and existing in the form of random dispersion during the actual operation of the wind and solar power station is classified as type two; the abnormal data caused by planned station maintenance, line failures and continuous energy abandonment and power rationing are classified as type three, among which type three has two forms of expression, one is accumulation in the form of horizontal lines, and the other form of expression is closely connected with normal data; the abnormal data caused by temporary station maintenance, deep participation of wind and solar power stations in power system dispatching and intermittent energy abandonment and power rationing factors are classified as type four.

7. The multi-scale fusion prediction method for wind and solar power generation in cold regions according to claim 1 is characterized in that: It also includes the identification and restoration of abnormal data in the collected wind and solar power station operation data. The specific workflow is as follows: The wind speed and irradiance are divided according to a certain value by using the longitudinal unilateral quartile method to obtain the power sequence in the meteorological interval, and the three quantiles of the power sequence in the meteorological interval are calculated. On this basis, the interquartile range is calculated, IQR = Q3-Q1 and the lower boundary L l =Q1-1.5×IQR. When the power in the meteorological interval is less than the lower boundary, it is regarded as abnormal data. The power is divided according to the percentage of installed capacity using the horizontal unilateral quartile method to obtain the meteorological sequence in the power interval. The three quantiles of the meteorological sequence in each power interval are calculated, and on this basis, the interquartile range and the upper boundary are calculated. L u =Q3+1.5×IQR, when the wind speed and irradiance within the power range are greater than the upper boundary, they are regarded as abnormal data.

8. The multi-scale fusion prediction method for wind and solar power generation in cold regions according to claim 1 is characterized in that: The method of using time series transfer to correct the wind speed / irradiance error of numerical weather forecast specifically includes the following steps: Sp1, NWP wind speed correction, the input variable set consists of the NWP wind speed at time t+1 and the measured wind speed sequence at time t, and the output variable set consists of the measured wind speed sequence at time t+1; NWP irradiance correction, the input variable set consists of the NWP irradiance at time t+1, the time label at time t+1 and the measured irradiance sequence at time t, and the output variable set consists of the measured irradiance sequence at time t+1; Sp2, 1m / s and 100W / m 2 The NWP wind speed and irradiance data are segmented for intervals, and then the input variable set and the output variable set are simultaneously divided into k interval segments; Sp3. Establish a mapping relationship between input variables and output variables in each interval, train wind speed and irradiance correction models, and store model parameters in each interval for future use in correcting NWP wind speed / irradiance; Sp4, based on the NWP wind speed / irradiance at time t+1, the model required for NWP correction is determined in real time, and the rolling iteration mode is used for correction. Within the correction time n, except for the first correction point, the correction value of the previous moment is substituted into the model input end as the measured value for NWP correction, and the NWP correction results under each correction time are obtained. Based on this, a database of NWP wind speed / irradiance correction results under different time lengths is established; Sp5. Based on the above-obtained correction result database, construct the power prediction input and output variable sets. For wind power prediction, the input variable set consists of the NWP corrected wind speed at time t+1 and the NWP wind direction sequence at time t+1, and the output variable set consists of the measured wind power sequence at time t+1; for photovoltaic power prediction, the input variable set consists of the NWP corrected irradiance at time t+1, the NWP temperature at time t+1, and the time label sequence at time t+1, and the output variable set consists of the measured photovoltaic power sequence at time t+1; Sp6. Divide each input variable set and output variable set by month based on time, establish a wind power / photovoltaic power generation prediction model in each month, and obtain wind power / photovoltaic power generation prediction results at different time scales.

9. The multi-scale fusion prediction method for wind and solar power generation in cold regions according to claim 1 is characterized in that: The wind and solar power generation power and load joint prediction based on deep learning includes the following steps: Adopt variable attention mechanism to extract key input information of wind and solar load; Based on multi-task learning and fully connected neural network, wind power prediction tasks and optical power prediction tasks are carried out simultaneously, and homoscedastic uncertainty is introduced to achieve automatic optimization of loss weights for different prediction tasks; Based on the long short-term memory neural network, a load forecasting model considering future wind and solar power generation information is established to predict the power of each wind farm and the total power of regional wind farms, the power of each photovoltaic power station and the total power of regional photovoltaic power stations.

10. The multi-scale fusion prediction method for wind and solar power generation in cold regions according to claim 9 is characterized in that: The specific steps of the wind and solar power generation power and load joint prediction based on deep learning are as follows: Construct the input variable set and output variable set required for wind power prediction and photovoltaic power prediction. The input variable set consists of the numerical weather forecast of each wind farm and each photovoltaic power station, including wind speed, irradiance and temperature. The output variable set consists of the power of each wind farm and each photovoltaic power station, and the total power of regional wind farms and regional photovoltaic power stations. Multi-task learning is implemented using hard parameter sharing. The hidden layer is shared among all tasks, and the output layer of each task is constructed separately. The total loss function of multi-task learning is Where L total is the total loss function of each task, w i is the weight of the i-th task, L i is the loss function of the i-th task. The Bayesian modeling idea is used to automatically obtain the optimal weight of each task. The loss functions of wind power prediction task and photovoltaic power prediction task are constructed based on homoscedastic uncertainty. A fully connected neural network is used to build a shared hidden layer in the joint prediction of wind power and photovoltaic power generation. The quantile regression algorithm is used to realize the uncertainty prediction of wind and solar power generation. The constructed wind-solar joint prediction model is used to predict the power of each wind farm and the total power of regional wind farms, the power of each photovoltaic power station and the total power of regional photovoltaic power stations.

Images