A renewable energy historical output fine reconstruction method

Abstract

The application discloses a renewable energy historical output fine reconstruction method, relates to the field of new energy power, and aims to solve the problems that low time resolution historical output data is difficult to accurately reflect the short-time fluctuation characteristics of renewable energy and that existing reconstruction methods are difficult to consider the adaptability of complex terrain and reconstruction accuracy; the method comprises the following steps: acquiring satellite remote sensing data, screening and constructing training samples; constructing an error estimation model, predicting a correction error field based on a linear interpolation benchmark field, and combining a linear interpolation method to obtain a comprehensive estimation model; constructing a bidirectional optical flow estimation model, weighting and synthesizing the optical flow information of output fields at previous and subsequent time points to obtain an optical flow output field; constructing a dynamic weight distribution model based on a terrain complexity index, outputting the fusion weight of the comprehensive estimation field and the optical flow output field, and obtaining a final reconstructed output field through consistency correction. The method realizes high time resolution fine reconstruction of renewable energy historical output data.

Description

Technical Field

[0001] This invention relates to the field of new energy power, specifically to a method for fine-grained reconstruction of historical output of renewable energy. Background Technology

[0002] In the planning and configuration of existing energy systems, satellite remote sensing data is typically used to obtain historical wind speed and solar irradiance data for the target area, and then further generates historical output sequences for wind and solar power. However, this type of data usually has low temporal resolution, mostly on the hourly level, making it difficult to accurately reflect the short-term fluctuation characteristics of renewable energy output. This can easily lead to problems with the actual operation of the planned installed capacity of wind power, solar power, energy storage, and other equipment in the energy system.

[0003] Chinese patent CN118229519A discloses a satellite sequence image interpolation method based on multimodal fusion optical flow estimation, which mainly involves interpolation and reconstruction of the image data itself to improve image reconstruction accuracy and visual continuity. This is fundamentally different from the technical solution of this invention, which addresses the need to improve the temporal resolution of historical output data of wind power and photovoltaic renewable energy. Summary of the Invention

[0004] To address the challenges of low temporal resolution historical output data failing to accurately reflect the short-term fluctuations of renewable energy, and the difficulty of existing reconstruction methods in balancing adaptability to complex terrain and reconstruction accuracy, this invention proposes a refined reconstruction method for renewable energy historical output data. This method achieves high temporal resolution refined reconstruction of renewable energy historical output data, thereby improving the accuracy and reliability of energy system configuration planning.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for refined reconstruction of historical renewable energy output, comprising the following steps: S1. Acquire satellite remote sensing data, filter and construct training samples; S2, construct an error estimation model, predict the correction error field based on the linear interpolation benchmark field, and obtain a comprehensive estimation model by combining the linear interpolation method; S3, construct a two-way optical flow estimation model, and weight and synthesize the optical flow information of the output force field at previous and subsequent time steps to obtain the optical flow output force field; S4 constructs a dynamic weight allocation model based on the terrain complexity index, outputs the fusion weights of the comprehensive estimation field and the optical flow output field, and obtains the final reconstructed output field after consistency correction.

[0006] In this technical solution, force field reconstruction results are constructed using convolutional neural network field estimation method and bidirectional optical flow estimation method, respectively. Furthermore, topographic feature parameters such as slope, elevation change rate, and topographic relief are introduced to construct a dynamic weight allocation model. The fusion ratio of the two reconstruction methods (convolutional neural network field estimation method and bidirectional optical flow estimation method) is adaptively adjusted according to different geographic grids and different topographic conditions to achieve automatic selection of the optimal reconstruction strategy under different regions.

[0007] Further, S1 includes: S11, acquire historical renewable energy satellite remote sensing data and geographic information data of the target area; S12, based on the physical quantity fluctuation index and weight, is weighted summation to obtain the comprehensive weather factor for that date; S13, sorted by comprehensive weather factors, takes the top a% high values ​​as training samples.

[0008] In this technical solution, the target area for new energy resource assessment or installed capacity optimization is determined, and then resource data within the target area is obtained through satellite remote sensing data; then, a comprehensive weather factor is calculated, and a corresponding training sample set is constructed based on the comprehensive weather factor; 'a' is a number greater than 0 and less than 100.

[0009] The training sample set mainly includes cloudy, windy, rainy, severe convection, sudden changes in wind speed, rapid changes in radiation, and other weather processes with distinct characteristics.

[0010] Furthermore, the construction of the linear interpolation reference field includes: By combining the time position ratio of the time layer to be reconstructed between two known observation times, the spatial force field of the previous time moment and the spatial force field of the next time moment are weighted and synthesized to obtain the linear interpolation reference field corresponding to the time layer to be reconstructed.

[0011] In this technical solution, the linear interpolation reference field is equal to the product of the output field at the previous known observation time and the output field at the time closest to the previous observation time, plus the ratio of the output field at the next known observation time to the output field at the time closest to the next observation time. The ratios closest to the previous observation time and the ratios closest to the next observation time are determined by the position of the time layer to be reconstructed between the two known observation times, and the sum of the two is 1.

[0012] Further, S2 includes: S21: Select multiple consecutive known observation times, use the data from the first and last two times as input and the data from the middle time as labels, and automatically construct a large number of fine samples by sliding along the time axis; S22, construct an error estimation model based on a convolutional neural network, output a corrected error field about the linear interpolation reference field, and superimpose it onto the linear interpolation reference field to obtain a comprehensive estimation field.

[0013] In this technical solution, a large number of fine samples are constructed using a self-supervised learning approach. Each fine sample includes the output force field at the previous observation time, the output force field at the next observation time, the intermediate true output force field, the linear interpolation reference field, the terrain feature field, the output force field corrected by the convolutional neural network, the output force field estimated by optical flow, and the final comprehensive estimate.

[0014] Further, S3 includes: S31, treat the output field of the previous moment and the next moment as a continuous two-dimensional field, and estimate the forward optical flow and the reverse optical flow; S32, the forward candidate output force field is obtained based on the forward optical flow, and the reverse candidate output force field is obtained based on the reverse optical flow; S33, the forward and backward candidate output fields are weighted according to the time position ratio to obtain the optical flow estimated output field.

[0015] In this technical solution, the two-way optical flow estimation model can estimate the spatial migration relationship of the output force field between two known observation times.

[0016] Further, S4 includes: Calculate the terrain complexity index; construct a dynamic weight allocation model, which includes an input layer, a hidden layer, and an output layer. The input layer receives terrain feature parameters, the comprehensive estimation model results, and the bidirectional optical flow estimation model results; the output layer outputs the fusion weights corresponding to the comprehensive estimation model and the fusion weights corresponding to the bidirectional optical flow estimation model, respectively.

[0017] In this technical solution, the hidden layer of the dynamic weight allocation model learns the nonlinear mapping relationship between terrain conditions, reconstruction results and final reconstruction error, thereby identifying the applicability of the two reconstruction methods under different terrain environments.

[0018] Furthermore, the input layer of the error estimation model receives the spatial output field at the previous known observation time, the spatial output field at the next known observation time, the spatial output field at the verification time, the linear interpolation reference field, and the terrain feature field; the hidden layer of the error estimation model contains multiple convolutional units, and normalization, nonlinear activation functions, and downsampling layers are set between the convolutional units.

[0019] Further, S33 includes: Based on forward optical flow, the output field from the previous moment is moved to the moment to be reconstructed, and based on reverse optical flow, the output field from the next moment is moved in the opposite direction to the moment to be reconstructed, thus obtaining the forward candidate output field and the reverse candidate output field, respectively.

[0020] Further, S4 includes: obtaining a comprehensive estimation result by outputting the fusion weights of the comprehensive estimation field and the optical flow output field. The comprehensive estimation result is the sum of the first weight term and the second weight term. The first weight term is the product of the comprehensive estimation model result and the corresponding weight, and the second weight term is the product of the bidirectional optical flow estimation model result and the corresponding weight.

[0021] Furthermore, the terrain complexity index is the sum of the slope weight term and the surrounding elevation difference weight term. The slope weight term is the product of the slope normalization index and the slope weight, and the surrounding elevation difference weight term is the product of the surrounding elevation difference normalization index and the surrounding elevation difference weight.

[0022] The present invention can bring the following beneficial effects: 1. By employing multiple technical means to comprehensively improve the historical temporal resolution of wind and solar power generation calculated from satellite remote sensing data, the accuracy of the data is significantly improved compared to conventional interpolation methods. This provides more accurate data for the optimization of renewable energy system configuration, thereby enhancing the stability of system operation. 2. It integrates convolutional neural network comprehensive estimation and optical flow estimation methods; it uses optical flow to accurately capture the smooth movement trajectory of clouds or airflows, and uses convolutional neural network comprehensive estimation methods to correct local abrupt changes; it effectively solves the problem of not being able to take into account both macroscopic continuous evolution and microscopic instantaneous fluctuations, and achieves accurate prediction of the spatiotemporal evolution of the entire field. 3. By incorporating terrain features as dynamic weights into the dynamic weight allocation model, the weights for convolutional neural network evaluation and optical flow estimation are automatically allocated based on the terrain complexity and output variation consistency of each geographic grid. This effectively overcomes the limitations of a single model and employs physical consistency correction to ensure that the output data is not only visually clear but also physically rigorous and usable, significantly improving the overall reconstruction accuracy. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the error iteration of the comprehensive valuation model of this invention.

[0024] Figure 2 This is a schematic diagram of the linear interpolation of the comprehensive valuation model of this invention.

[0025] Figure 3 This is a flowchart of the training process for the dynamic weight allocation model of this invention.

[0026] Figure 4 This is a schematic diagram comparing the effects of the interpolation method and the comprehensive evaluation method of this invention (wind power capacity factor).

[0027] Figure 5 This is a schematic diagram comparing the effects of the interpolation method and the comprehensive evaluation method of this invention (photovoltaic capacity factor). Detailed Implementation

[0028] Example 1:

[0029] In existing energy system configuration planning, it is typically necessary to first obtain historical wind speed, solar irradiance, temperature, and meteorological resource data for the target area using satellite remote sensing data. Then, historical output sequences for wind power and photovoltaics are obtained through power conversion models. These are used as basic data and combined with load demand data to analyze and determine the optimal installed capacity of various equipment in the energy system. However, the temporal resolution of satellite remote sensing data is usually low, mostly at hourly accuracy, which cannot adapt to the rapid fluctuations in renewable energy generation. Therefore, it can cause significant errors in energy system configuration planning, thus affecting the normal operation of the energy system. Therefore, it is necessary to improve the temporal resolution of historical output for wind power and photovoltaics. To address the problem that low temporal resolution historical output data cannot accurately reflect the short-term fluctuation characteristics of renewable energy, and that existing reconstruction methods cannot balance adaptability to complex terrain and reconstruction accuracy, a refined reconstruction method for historical output of renewable energy is proposed to achieve high temporal resolution refined reconstruction of historical output data. This method includes the following steps.

[0030] Step S1: Acquire satellite remote sensing data, filter and construct training samples; this step mainly includes the following sub-steps.

[0031] Step S11: Obtain historical renewable energy satellite remote sensing data and geographic information data for the target area.

[0032] For step S11, firstly, the target area for new energy resource assessment or installed capacity optimization is determined; then, meteorological resource data such as historical wind speed, solar irradiance, and temperature for each geographic grid within the target area are obtained through satellite remote sensing data, and the corresponding wind power and photovoltaic power generation capacity factors are calculated. Simultaneously, geographic elevation data of the target area is acquired, and based on the elevation data, the slope, aspect, and topographic relief characteristics of each geographic grid are further calculated.

[0033] Step S12: Perform a weighted summation based on the physical quantity fluctuation index and weights to obtain the comprehensive weather factor for that date.

[0034] For step S12, in order to improve the learning ability of the subsequent model for drastic weather change processes, a comprehensive weather factor parameter is constructed to evaluate the intensity and complexity of daily meteorological process fluctuations.

[0035] For each historical date, the daily variance of each core weather physical quantity (wind speed, temperature, and solar irradiance) is calculated for that date. The daily variance reflects the fluctuation range of the physical quantity on that date. The normalized fluctuation indicators of each physical quantity are weighted and summed according to a set ratio to obtain the comprehensive weather factor corresponding to that date.

[0036] Specifically, the comprehensive weather factor equals the normalized wind speed volatility index multiplied by the wind speed weight + the normalized temperature volatility index multiplied by the temperature weight + the normalized solar irradiance volatility index multiplied by the solar irradiance weight. The larger the comprehensive weather factor, the more intense the meteorological process on that date, the more complex the changes in renewable energy output, and the more suitable it is as a training sample for subsequent models.

[0037] Step S13: Sort by comprehensive weather factors and take the top a% high values ​​as training samples; a is a number greater than 0 and less than 100.

[0038] For step S13, after calculating the comprehensive weather factor in step S12, the comprehensive weather factors of all historical dates are sorted. A subset of dates with higher comprehensive weather factors (in this embodiment, the first quarter of the sample) are selected as the training sample set. This training sample set mainly includes cloudy, windy, rainy, severe convection, sudden wind speed changes, rapid changes in radiation, and other weather processes with distinct characteristics. In this type of sample, the linear interpolation error is relatively large, and the characteristics of wind and solar power output changes are more prominent. However, it is suitable for use in model pre-training, allowing the model to prioritize learning key features such as rapid changes in renewable energy power output, local abrupt changes, and spatial migration.

[0039] Step S2: Construct an error estimation model, predict the correction error field based on the linear interpolation benchmark field, and obtain a comprehensive estimation model by combining the linear interpolation method.

[0040] In step S2, the construction of the linear interpolation reference field includes the following process: combining the time position ratio of the time layer to be reconstructed between two known observation times, the spatial force field of the previous time and the spatial force field of the next time are weighted and synthesized to obtain the linear interpolation reference field corresponding to the time layer to be reconstructed.

[0041] For each time layer to be reconstructed (e.g., estimating wind and solar power output data at time t1 based on data at times t0 and t2), a linear difference reference field is constructed based on the spatial power output fields at the two known observation times before and after it (e.g., based on times t0 and t2).

[0042] The linear interpolation reference field equals the output field at the previous known observation time multiplied by the proportion closest to the previous time, plus the output field at the next known observation time multiplied by the proportion closest to the next time. The proportions closest to the previous time and the proportion closest to the next time are determined by the position of the time layer to be reconstructed between the two known observation times, and their sum is 1.

[0043] A linear interpolation reference field can describe the basic estimation results when the output force changes smoothly over time. For periods of gradual change, this result can serve as an effective approximation; however, for periods of rapid cloud movement, sudden wind speed changes, or significant topographic disturbances, the result typically contains large errors. Therefore, this invention does not directly use the linear interpolation reference field as the final result, but rather uses it as the basis for subsequent convolutional neural network error correction and optical flow estimation fusion.

[0044] Step S2 mainly includes the following sub-steps.

[0045] Step S21: Select multiple consecutive known observation times, use the data from the first and last two times as input and the data from the middle time as labels, and automatically construct a large number of fine samples by sliding along the time axis.

[0046] More specifically, a self-supervised learning approach is used to construct training samples. Multiple consecutive known observation times are selected from historical low temporal resolution data. Satellite remote sensing data at two consecutive observation times (e.g., t0 and t2) are used as input, while satellite remote sensing data at a known observation time in between (e.g., t1) is used as a supervisory reference. The model reconstructs the intermediate observation time value based on the t0 and t2 observation times and compares the reconstructed result with the true intermediate observation time (t1) value.

[0047] By sliding this sampling method along the time axis, a large number of fine samples can be automatically constructed from years of historical data without additional manual annotation. Each fine sample includes the output force field at the previous observation time, the output force field at the next observation time, the intermediate true output force field, the linear interpolation reference field, the terrain feature field, the output force field corrected by the convolutional neural network, the output force field estimated by optical flow, and the final comprehensive estimate.

[0048] Step S22: Construct an error estimation model based on a convolutional neural network, output a corrected error field about the linear interpolation reference field, and superimpose it onto the linear interpolation reference field to obtain a comprehensive estimation field.

[0049] For step S22, the constructed error estimation model can mainly evaluate the spatial error between the linear difference reference field and the actual output field. The core of this error estimation model is to combine terrain features to perform nonlinear correction on the smoothing result of the linear difference. Then, the error estimation model is combined with the linear interpolation method to obtain a comprehensive estimation model based on a convolutional neural network. The comprehensive estimation field is obtained based on the comprehensive estimation model.

[0050] The input layer of the error estimation model receives the spatial power output field at the previous known observation time (t0), the spatial power output field at the next known observation time (t2), the spatial power output field at the verification time (t1), the linear interpolation reference field, and the terrain feature field. Among them, the terrain feature field serves as an auxiliary input channel, which can guide the convolutional neural network to learn the constraint effect of complex terrain on the spatial distribution of renewable energy power output.

[0051] The hidden layer of the error estimation model contains multiple convolutional units, with normalization, nonlinear activation functions, and downsampling layers (such as max pooling or average pooling) between the convolutional units. This can extract spatial features from the input data and reduce the computational dimensionality. The hidden layer can adopt a residual connection structure or a dense connection structure to enhance the model's ability to learn from local terrain abrupt changes and nonlinear changes in output.

[0052] The output layer of the error estimation model outputs a two-dimensional correction error field with the same spatial resolution as the input. This correction error field represents the potential deviation of the linear interpolation result from the actual power output state at each geographic grid. The error estimation model does not directly output the final power output field; instead, it outputs the correction error field of the linear interpolation reference field. This correction error field represents the potential deviation of the linear interpolation result from the actual power output state at each geographic grid.

[0053] The comprehensive valuation field equals the linear interpolation reference field plus the correction error field. The linear interpolation reference field can preserve the basic time trend, while the correction error field can compensate for errors caused by nonlinear changes, local abrupt changes, terrain shading, cloud boundary changes, and local wind field disturbances.

[0054] refer to Figure 2 During the model training phase, the corrected error field f output by the error estimation model based on the convolutional neural network is superimposed with the linear interpolation result q to obtain the estimated output force field F', and the loss function, i.e., the error e, is calculated using the actual observed output force field at that moment as the true value F.

[0055] Building upon this, the difference between the estimated output field and the actual observed field is minimized using the backpropagation algorithm, thereby iteratively updating the network weights. This allows the model to gradually learn how to accurately compensate for the deviations of linear interpolation based on terrain and input features. For detailed steps, please refer to [reference needed]. Figure 1 .

[0056] Step S3: Construct a bidirectional optical flow estimation model, and synthesize the optical flow information of the output force field at previous and subsequent time points by weighting to obtain the optical flow output force field. This step mainly includes the following sub-steps.

[0057] Step S31: Treat the output field at the previous and next time moments as a continuous two-dimensional field, and estimate the forward and reverse optical flows. The output field is the spatial distribution field of renewable energy-related parameters output by the satellite remote sensing sensor.

[0058] Step S32: Obtain the forward candidate output force field based on the forward optical flow, and obtain the reverse candidate output force field based on the reverse optical flow.

[0059] Step S33: Combine the forward and reverse candidate output fields according to the time position ratio to obtain the optical flow estimated output field.

[0060] The bidirectional optical flow estimation model of this invention can estimate the spatial migration relationship of the output force field between two known observation times.

[0061] First, the spatial output force field at the previous known observation time (t0) and the spatial output force field at the next known observation time (t2) are regarded as continuously changing two-dimensional fields. Then, the forward optical flow from the previous known observation time to the next known observation time is estimated to describe the spatial displacement of the output force field as it evolves forward with time. At the same time, the reverse optical flow from the next known observation time to the previous known observation time is estimated to describe the spatial displacement of the output force field as it moves backward.

[0062] Based on the forward optical flow (which can be understood as analyzing the flow vector field of the optical flow based on data from time t0 to time t2), the output field at the previous known observation time is moved to the position corresponding to the time layer to be reconstructed, thus obtaining the candidate output field of the forward optical flow; based on the reverse optical flow (obtaining the reverse optical flow vector field based on data from time t2 and time t0), the output field at the next known observation time is moved in the reverse direction to the position corresponding to the time layer to be reconstructed, thus obtaining the candidate output field of the reverse optical flow.

[0063] Based on forward optical flow, the output field from the previous moment is moved to the moment to be reconstructed, and based on reverse optical flow, the output field from the next moment is moved in the opposite direction to the moment to be reconstructed, thus obtaining the forward candidate output field and the reverse candidate output field, respectively.

[0064] Based on the temporal position ratio of the time layer to be reconstructed between two known observation times, the forward optical flow candidate output fields and the reverse optical flow candidate output fields are weighted and synthesized. That is, the estimated optical flow output field equals the forward optical flow candidate output field multiplied by the proportion closer to the previous time step + the reverse optical flow candidate output field multiplied by the proportion closer to the next time step.

[0065] Step S4: Construct a dynamic weight allocation model based on the terrain complexity index, output the fusion weights of the comprehensive estimated field and the optical flow output field, and obtain the final reconstructed output field after consistency correction.

[0066] The terrain complexity index is the sum of the slope weight term and the surrounding elevation difference weight term. The slope weight term is the product of the slope normalization index and the slope weight, and the surrounding elevation difference weight term is the product of the surrounding elevation difference normalization index and the surrounding elevation difference weight.

[0067] Due to different terrain conditions, the comprehensive estimation model is more accurate in some areas, while the optical flow estimation method is more accurate in others.

[0068] To determine the applicability of the convolutional neural network-corrected output force field and the optical flow-estimated output force field under different geographic grids, this invention further calculates the terrain complexity index.

[0069] The terrain complexity index is determined by slope and the difference in elevation between the surrounding areas (terrain undulation). A steeper slope, a greater difference in elevation between the surrounding areas, and stronger terrain undulation indicate a more complex terrain; conversely, a gentler slope, a smaller difference in elevation between the surrounding areas, and weaker terrain undulation indicate a more level terrain. This can be expressed by the following formula: The terrain complexity index equals the slope normalization index multiplied by the slope weight, plus the surrounding elevation difference normalization index multiplied by the surrounding elevation difference weight.

[0070] For step S4, refer to Figure 3 It includes: calculating the terrain complexity index; and constructing a dynamic weight allocation model, which includes an input layer, a hidden layer, and an output layer.

[0071] The input layer receives the terrain feature parameters corresponding to the target geographic grid, the results of the comprehensive estimation model, and the results of the two-way optical flow estimation model; the terrain feature parameters include slope, elevation change rate or terrain relief, and other indicators that can reflect the complexity of the regional terrain.

[0072] The hidden layer can learn the nonlinear mapping relationship between terrain conditions, reconstruction results and final reconstruction error, thereby identifying the applicability of the two reconstruction methods under different terrain environments.

[0073] The output layer outputs the fusion weights corresponding to the comprehensive estimation model and the bidirectional optical flow estimation model, respectively. Specifically, the output layer contains two output nodes, which output the fusion weights corresponding to the estimation methods of the comprehensive estimation model and the optical flow estimation method, respectively. Both fusion weights are non-negative, and their sum is 1. Weighted fusion is then performed to obtain the final comprehensive estimation result.

[0074] By merging the weights of the comprehensive estimation field and the optical flow output field, a comprehensive estimation result is obtained. The comprehensive estimation result is the sum of the first weight term and the second weight term. The first weight term is the product of the comprehensive estimation model result and the corresponding weight, and the second weight term is the product of the bidirectional optical flow estimation model result and the corresponding weight.

[0075] The dynamic weight allocation model constructed in step S4 can determine the relative weights of the comprehensive estimated output field and the optical flow estimated output field for each geographic grid and each time layer to be reconstructed.

[0076] During the training of the dynamic weight allocation model, the final comprehensive estimation result is compared with the true result, and the reconstruction error is calculated. Based on the reconstruction error, the parameters of the dynamic weight allocation model are updated through backpropagation, enabling the model to automatically learn the optimal weight ratio of the two reconstruction methods under different terrain conditions. The specific calculation process is as follows: Figure 3 As shown.

[0077] The fusion process is executed independently on each geographic grid, so different geographic grids can have different fusion weights within the same time frame to be reconstructed. After training, when the terrain of the target area is complex and local variations are significant, the model tends to increase the fusion weight of the estimation method in the comprehensive estimation model; when the terrain is relatively flat and the output changes have strong spatial continuity, the model tends to increase the fusion weight of the optical flow estimation method, thereby achieving adaptive adjustment of the reconstruction strategy under different geographic grids and improving the accuracy of time-refined reconstruction of historical output data.

[0078] After obtaining the final comprehensive estimation result (i.e. the final reconstructed force field), a consistency correction is performed, specifically a physical consistency correction.

[0079] For photovoltaic (PV) output data, the first step is to determine whether the target time period falls within an effective solar irradiance period. If it falls within a period without solar irradiance, the PV output is corrected to zero. If it falls within an effective solar irradiance period, the PV output range is limited based on the upper limit of solar irradiance, PV installed capacity, and capacity factor boundaries. For time periods near sunrise and sunset, the rate of change in PV output is limited to avoid unreasonable abrupt changes.

[0080] For wind power output data, first determine whether the reconstructed result is less than zero or exceeds the rated output boundary; if it is less than zero, correct it to zero; if it exceeds the rated boundary, correct it to the rated boundary. Secondly, combine the wind turbine power curve to perform a consistency check on the cut-in, rated, and cut-out operating ranges.

[0081] For spatial anomalies, if the reconstruction output of a geographic grid deviates significantly from the surrounding grids, and this deviation cannot be explained by terrain elevation, slope, aspect, or occlusion features, then the grid is marked as an anomaly and corrected by combining the results of the surrounding grids and adjacent time layers.

[0082] After completing steps S1 to S4 above, a historical dataset is generated that can be used to optimize installed capacity and the optimal configuration result of the energy system is solved.

[0083] Specifically, the original power output fields at all known observation times and the final reconstructed power output fields at all time layers to be reconstructed are arranged in chronological order to form a historical power output dataset of renewable energy at the target time resolution.

[0084] The dataset includes at least the overall wind power capacity factor sequence of the target area, the overall photovoltaic capacity factor sequence of the target area, the dynamic weight allocation results, and the error evaluation results.

[0085] Based on this, the high time-resolution historical output data of wind power and photovoltaic power are input into the energy system configuration optimization model. Combined with the installation cost of energy storage and other equipment, the installed capacity optimization model can aim at minimizing the total system cost. Considering battery charging and discharging constraints, power generation and consumption balance constraints, and transmission channel constraints, the optimal wind power installed capacity, photovoltaic installed capacity, energy storage installed power, and energy storage installed capacity in the energy system are obtained by using the mixed integer linear programming method.

[0086] The above-described technical solution in this embodiment can bring about the following technical effects.

[0087] First, by employing multiple technical means to comprehensively improve the historical temporal resolution of wind and solar power generation calculated from satellite remote sensing data, the accuracy of the data is significantly improved compared to conventional interpolation methods. This provides more accurate data for optimizing the configuration of renewable energy systems, thereby enhancing the stability of system operation.

[0088] Second, it integrates convolutional neural network (CNN) estimation and optical flow estimation methods. Optical flow is used to accurately capture the smooth movement trajectory of clouds or air currents, while the CNN estimation method corrects for local abrupt changes. This mechanism effectively solves the challenge of simultaneously capturing macroscopic continuous evolution and microscopic instantaneous fluctuations, achieving accurate prediction of the entire spatiotemporal evolution.

[0089] Third, terrain features are incorporated as dynamic weights into the dynamic weight allocation model. Based on the terrain complexity and output variation consistency of each geographic grid, the weights for convolutional neural network evaluation and optical flow estimation are automatically allocated, which effectively overcomes the limitations of a single model. Furthermore, physical consistency correction is adopted to ensure that the output data is not only visually clear but also physically rigorous and usable, significantly improving the overall reconstruction accuracy.

[0090] Example 2:

[0091] Based on Example 1, taking a region in Northeast my country as an example, a large number of measuring devices were installed in this area, recording a large amount of historical renewable energy output observation data, with a maximum time resolution of 0.5 hours (in reality, the current mainstream method for obtaining historical wind and solar power output data in the study area through satellite remote sensing data has a time accuracy of 1 hour; to highlight the role of this invention, the actual equipment measurement results are used as a comparative case for analysis). Based on this, typical days were selected for comparison. The effectiveness of the method of this invention was verified, and the results were compared with those of traditional interpolation methods. Specific comparison results of wind power and solar power output are as follows: Figure 4 and Figure 5 As shown.

[0092] The above analysis shows that the precision of the refined results obtained using the comprehensive evaluation method of this invention is significantly higher than that of the linear interpolation method. Based on this, a mixed-integer linear programming method is used to solve and analyze the optimal configuration of the energy system in this region, as shown in Table 1. Table 1. Comparison of optimal energy system configuration results under different methods

[0093] Analysis revealed that energy system configurations obtained using low-resolution wind and solar power output data have relatively low redundancy, potentially leading to downtime in the future. The method described in this invention can effectively refine historical wind and solar power output data, yielding results that more closely reflect reality and improving the safety of energy system operation to some extent.

Claims

1. A method for refined reconstruction of historical renewable energy output, characterized in that, Includes the following steps: S1. Acquire satellite remote sensing data, filter and construct training samples; S2, construct a comprehensive valuation model, predict the correction error field, and combine it with the linear interpolation benchmark field to obtain the comprehensive valuation field; S3, construct a two-way optical flow estimation model, and weight and synthesize the optical flow information of the output force field at previous and subsequent time steps to obtain the optical flow output force field; S4 constructs a dynamic weight allocation model based on the terrain complexity index, outputs the fusion weights of the comprehensive estimation field and the optical flow output field, and obtains the final reconstructed output field after consistency correction.

2. The method for refined reconstruction of historical renewable energy output according to claim 1, characterized in that, S1 includes: S11, acquire historical renewable energy satellite remote sensing data and geographic information data of the target area; S12, based on the physical quantity fluctuation index and weights, is weighted summation to obtain the comprehensive weather factor for that date; S13, sorted by comprehensive weather factors, takes the top a% high values ​​as training samples.

3. A method for refined reconstruction of historical renewable energy output according to claim 1 or 2, characterized in that, The construction of the linear interpolation reference field includes: By combining the time position ratio of the time layer to be reconstructed between two known observation times, the spatial force field of the previous time moment and the spatial force field of the next time moment are weighted and synthesized to obtain the linear interpolation reference field corresponding to the time layer to be reconstructed.

4. The method for refined reconstruction of historical renewable energy output according to claim 3, characterized in that, S2 includes: S21: Select multiple consecutive known observation times, use the data from the first and last two times as input and the data from the middle time as labels, and automatically construct a large number of fine samples by sliding along the time axis; S22. Construct a comprehensive estimation model based on a convolutional neural network, output a correction error field for the linear interpolation reference field, and superimpose it onto the linear interpolation reference field to obtain the comprehensive estimation field.

5. A method for refined reconstruction of historical renewable energy output according to claim 1, 2, or 4, characterized in that, S3 includes: S31, treat the output field of the previous moment and the next moment as a continuous two-dimensional field, and estimate the forward optical flow and the reverse optical flow; S32, the forward candidate output force field is obtained based on the forward optical flow, and the reverse candidate output force field is obtained based on the reverse optical flow; S33, the forward and backward candidate output fields are weighted according to the time position ratio to obtain the optical flow estimated output field.

6. The method for refined reconstruction of historical renewable energy output according to claim 5, characterized in that, S5 includes: Calculate the terrain complexity index; construct a dynamic weight allocation model, which includes an input layer, a hidden layer, and an output layer. The input layer receives terrain feature parameters, the comprehensive estimation model results, and the bidirectional optical flow estimation model results; the output layer outputs the fusion weights corresponding to the comprehensive estimation model and the fusion weights corresponding to the bidirectional optical flow estimation model, respectively.

7. A method for refined reconstruction of historical renewable energy output according to claim 1 or 4, characterized in that, The input layer of the comprehensive estimation model receives the spatial force field at the previous known observation time, the spatial force field at the next known observation time, the spatial force field at intermediate time, the linear interpolation reference field, and the terrain feature field; the hidden layer of the comprehensive estimation model contains multiple convolutional units, with normalization, nonlinear activation functions, and downsampling layers set between the convolutional units.

8. The method for refined reconstruction of historical renewable energy output according to claim 5, characterized in that, S33 includes: Based on forward optical flow, the output field from the previous moment is moved to the moment to be reconstructed, and based on reverse optical flow, the output field from the next moment is moved in the opposite direction to the moment to be reconstructed, thus obtaining the forward candidate output field and the reverse candidate output field, respectively.

9. A method for refined reconstruction of historical renewable energy output according to claim 1 or 6, characterized in that, S4 includes: obtaining a comprehensive estimation result by fusion weights of the output comprehensive estimation field and the optical flow output field. The comprehensive estimation result is the sum of the first weight term and the second weight term. The first weight term is the product of the comprehensive estimation model result and the corresponding weight, and the second weight term is the product of the bidirectional optical flow estimation model result and the corresponding weight.

10. The method for refined reconstruction of historical renewable energy output according to claim 6, characterized in that, The terrain complexity index is the sum of the slope weight term and the surrounding elevation difference weight term. The slope weight term is the product of the slope normalization index and the slope weight, and the surrounding elevation difference weight term is the product of the surrounding elevation difference normalization index and the surrounding elevation difference weight.

Images