Interval generation and probability correction method and system for new energy power prediction

Abstract

The invention discloses an interval generation and probability correction method and system for new energy power prediction, and belongs to the technical field of power system operation and control, and the method comprises the steps: generating a multi-dimensional environment feature code through numerical weather forecast data, and calling a basic point prediction model to obtain a point prediction result and basic probability distribution; inputting the multi-dimensional environment feature code and the point prediction result into a dual-channel dynamic interval generator to form a preliminary prediction interval, calculating a short-term error sequence based on the actual power of new energy power generation and historical prediction data, analyzing the trend characteristics of the short-term error sequence, and correcting the basic probability distribution according to the trend characteristics; and extracting a probability verification signal from the corrected probability distribution, feeding back the probability verification signal to an interval generator, carrying out optimization adjustment on the preliminary prediction interval, and outputting an optimized prediction interval. According to the technical scheme, the preliminary interval is generated by adopting a dual-channel mechanism, and feedback correction and closed-loop optimization are performed in combination with the short-term error trend, so that the adaptive capacity, accuracy and reliability of the prediction interval can be improved.

Description

Technical Field

[0001] This invention relates to the field of power system operation and control technology, and in particular to a method and system for interval generation and probability correction of new energy power prediction. Background Technology

[0002] New energy power generation, especially wind and solar power, is playing an increasingly important role in the global energy mix. The intermittency, volatility, and randomness of its output pose significant challenges to the safe, stable operation and economic dispatch of power systems. To ensure a balance between power supply and demand, accurate forecasting of future power generation from new energy power plants is crucial. Interval forecasting, as an important supplement to point forecasting, quantifies the uncertainty of forecasts by providing the potential range of future power fluctuations. This plays a vital role in enabling the power grid to formulate reasonable reserve capacity, optimize dispatch schemes, and assess operational risks.

[0003] Among related technologies, Chinese invention patent CN114492964B discloses a method for ultra-short-term probabilistic prediction of photovoltaic power based on wavelet decomposition and optimized deep belief network. This method includes using wavelet decomposition to perform a three-level decomposition of the original data to obtain sub-sequences; using a deep belief network to establish power prediction models for each sub-sequence; using a genetic algorithm to find the optimal initial parameters for the models; and weighting the prediction results from each sub-model to obtain the final prediction result. Secondly, based on the constructed photovoltaic power point prediction model, the point prediction error sequence is analyzed, and an ultra-short-term probabilistic prediction model for photovoltaic power based on a t-distribution model is constructed to calculate the prediction interval.

[0004] Regarding the aforementioned technologies, the inventors believe that the above-mentioned solutions have problems in terms of uncertainty quantification. First, they rely too heavily on historical error statistics, failing to dynamically capture the variability caused by sudden weather changes or model drift, resulting in insufficient flexibility in the prediction interval width and potential inaccuracies. Second, the uncertainty differentiation and management are inadequate, failing to finely distinguish between the model's inherent uncertainty and external meteorological risks, causing the interval to fail to fully reflect various impacts. Third, the error analysis is one-off or periodic, unable to monitor online and dynamically adjust the interval position and width using short-term error trends, thus failing to correct persistent systematic biases in a timely manner, reducing the reliability and calibrability of the prediction. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method and system for generating and probabilistically correcting intervals for new energy power prediction. The method employs a dual-channel mechanism to generate an initial interval and combines it with short-term error trends for feedback correction and closed-loop optimization, thereby improving the adaptability, accuracy, and reliability of the prediction interval.

[0006] The above objectives can be achieved through the following approach:

[0007] A method and system for interval generation and probability correction of new energy power prediction includes: acquiring numerical weather forecast data and performing dynamic stability analysis based on the numerical weather forecast data to generate a multi-dimensional environmental feature code; calling a preset baseline point prediction model based on the numerical weather forecast data to obtain point prediction results and a baseline probability distribution; inputting the multi-dimensional environmental feature code and the point prediction results into a dual-channel dynamic interval generator to generate a preliminary prediction interval; collecting actual power data of new energy power generation and corresponding historical prediction data, calculating a short-term error sequence, performing trend analysis on the short-term error sequence to obtain short-term error trend characteristics; adjusting the baseline probability distribution using the short-term error trend characteristics to generate a corrected probability distribution; extracting a probability verification signal from the corrected probability distribution and feeding it back to the dual-channel dynamic interval generator; optimizing and adjusting the preliminary prediction interval based on the probability verification signal to generate and output an optimized prediction interval.

[0008] Optionally, generating the multidimensional environmental feature code includes: acquiring numerical weather forecast data containing wind speed, irradiance, and cloud change rate; calculating the fluctuation amplitude index and the trend index for a future period based on the numerical weather forecast data; and generating the multidimensional environmental feature code by analyzing and calculating the fluctuation amplitude index and the trend index.

[0009] Optionally, obtaining the point prediction results and the basic probability distribution includes: inputting the numerical weather forecast data into the basic point prediction model, extracting meteorological correlation features and time correlation features and performing dynamic weight fusion to obtain a fused feature vector; performing power trend extrapolation based on the fused feature vector to obtain the point prediction results; quantifying and estimating the uncertainty in the power trend extrapolation process, and forming a basic probability distribution through probability diffusion and confidence propagation mechanisms.

[0010] Optionally, generating the preliminary prediction interval includes: in the dual-channel dynamic interval generator, performing a difference analysis on the point prediction results at the same time based on multiple sub-models in the basic point prediction model to generate a prediction confidence signal; performing an environmental risk assessment based on the multi-dimensional environmental feature code to generate an environmental risk signal; fusing the prediction confidence signal and the environmental risk signal to obtain a fusion result, and constructing a dynamic interval based on the fusion result and the point prediction results to generate the preliminary prediction interval.

[0011] Optionally, obtaining the short-term error trend features includes: calculating the difference between the actual power and the historical predicted power point by point to form a short-term error sequence arranged in chronological order; performing local aggregation analysis within a sliding window on the short-term error sequence to identify the direction and rate of change of the error in adjacent time periods; and constructing a trend description vector reflecting the continuity and inertia of error evolution based on the direction and rate of change, as the short-term error trend features.

[0012] Optionally, generating the corrected probability distribution includes: identifying the same direction of continuous prediction errors in the short-term error sequence, generating systematic deviation direction and intensity characteristics; determining the distribution shift and distribution shape adjustment based on the systematic deviation direction and intensity characteristics; and adjusting the basic probability distribution using the distribution shift and distribution shape adjustment to generate the corrected probability distribution.

[0013] Optionally, adjusting the basic probability distribution using the distribution shift and the distribution shape adjustment includes: shifting the overall position of the basic probability distribution based on the distribution shift to compensate for the systematic deviation; after completing the overall position shift, reshaping the dispersion and skewness characteristics of the basic probability distribution according to the distribution shape adjustment; introducing an adaptive smoothing mechanism during the adjustment process to constrain overcorrection caused by local drastic fluctuations, and obtaining the corrected probability distribution.

[0014] Optionally, generating and outputting the optimized prediction interval includes: extracting a probability verification signal from the corrected probability distribution; parsing the probability verification signal to obtain an interval shift guidance and an interval width adjustment guidance; performing a secondary fusion of the interval shift guidance and the interval width adjustment guidance with the fusion result to obtain a secondary fusion result; and adjusting the position and width of the initial prediction interval based on the secondary fusion result to generate the optimized prediction interval.

[0015] Optionally, extracting the probability verification signal from the corrected probability distribution includes: calculating the overall offset of the corrected probability distribution relative to the basic probability distribution; generating the probability verification signal based on the overall offset and the systematic deviation direction and intensity characteristics; and sending the probability verification signal to the dual-channel dynamic range generator.

[0016] Based on the same inventive concept, this invention also provides a system for interval generation and probability correction of new energy power prediction. The system includes: a forecast feature generation module for acquiring numerical weather forecast data and performing dynamic stability analysis based on the numerical weather forecast data to generate a multi-dimensional environmental feature code; a point prediction distribution module for calling a preset basic point prediction model to obtain point prediction results and a basic probability distribution; a preliminary interval generation module for inputting the multi-dimensional environmental feature code and the point prediction results into a dual-channel dynamic interval generator to generate a preliminary prediction interval; an error trend analysis module for collecting actual power data of new energy power generation and corresponding historical prediction data, calculating a short-term error sequence, performing trend analysis on the short-term error sequence, and obtaining short-term error trend characteristics; a probability distribution correction module for adjusting the basic probability distribution using the short-term error trend characteristics to generate a corrected probability distribution; and an interval optimization output module for extracting a probability verification signal from the corrected probability distribution and feeding it back to the dual-channel dynamic interval generator, optimizing the preliminary prediction interval based on the probability verification signal, and generating and outputting an optimized prediction interval.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] 1. This invention, by constructing a dual-channel fusion mechanism of internal model confidence and external environmental risk, can dynamically generate a preliminary prediction interval that matches future weather conditions and model status. This allows the prediction interval width to be adaptively adjusted, narrowing during stable weather to provide more accurate decision support, and widening during periods of drastic weather fluctuations to ensure sufficient risk coverage, thereby improving the reliability and scenario adaptability of the prediction interval.

[0019] 2. This invention designs a feedback correction loop based on short-term error sequence analysis, which can capture and quantify the systematic or inertial deviations of the prediction model under the current operating conditions in real time. By utilizing this error trend characteristic to adjust the basic probability distribution, online self-correction of prediction uncertainty is achieved, effectively compensating for short-term prediction deviations caused by limitations in training data or operating condition drift, and improving the overall accuracy of prediction.

[0020] 3. This invention establishes a closed-loop feedback system from probability distribution correction to interval optimization. By extracting a probability verification signal from the corrected probability distribution and feeding it back to adjust the initial prediction interval, a secondary fine-tuning of the interval position and width is achieved. This closed-loop mechanism ensures that the final output prediction interval comprehensively reflects multi-dimensional information such as the model, environment, and recent actual performance, making it closer to reality.

[0021] 4. The method and system provided by this invention combine feedforward-based dynamic interval generation with feedback-based real-time probability correction to achieve a more comprehensive and refined quantification and management of prediction uncertainty. Compared with traditional methods, the optimized prediction interval output has higher confidence and accuracy, providing higher-quality decision-making basis for power system scheduling plans, reserve capacity allocation, and market transactions, thereby enhancing the security and economy of the power grid in accepting a high proportion of new energy sources.

[0022] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a method for interval generation and probability correction of new energy power prediction according to an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of short-term error sequence and trend analysis according to an embodiment of the present invention.

[0026] Figure 3 This is a comparison diagram of the basic probability distribution and the corrected probability distribution in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the structure of a new energy power prediction interval generation and probability correction system according to an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Reference Figure 1One embodiment of the present invention proposes a method for generating and probabilistically correcting the range of new energy power prediction. The method adopts a dual-channel mechanism to generate the initial range and combines it with the short-term error trend for feedback correction and closed-loop optimization, which can improve the adaptability, accuracy and reliability of the prediction range.

[0030] The method described in this embodiment specifically includes:

[0031] Numerical weather forecast data is acquired, and dynamic stability analysis is performed based on the numerical weather forecast data to generate a multidimensional environmental feature code.

[0032] Optionally, the generation of the multidimensional environmental feature code includes:

[0033] Acquire numerical weather forecast data including wind speed, irradiance, and cloud cover rate of change;

[0034] Based on the numerical weather forecast data, the fluctuation range index and the trend index for the future period are calculated.

[0035] Based on the fluctuation amplitude index and the change trend index, a multidimensional environmental feature code is generated through analysis and calculation.

[0036] Specifically, the process begins by acquiring numerical weather prediction data for the next 24 to 72 hours. This data is in time series format, typically with a time resolution of 15 minutes or 1 hour. The focus is on extracting variables directly related to renewable energy power generation, including wind speed, irradiance, and cloud cover rate of change. The subsequent index calculation process involves analyzing each extracted meteorological variable using a sliding window approach in the prediction time domain. The sliding window size is a key parameter, usually set to 1 to 3 hours to capture short-term meteorological dynamics. Within each window, two types of core indicators are calculated in parallel. The first type is the volatility index, which measures the instability or oscillation intensity of meteorological factors in the short term. This volatility index is typically obtained by calculating the standard deviation of the data series within the window. For example, for a wind speed series, the volatility index is calculated as follows: ,have:

[0037] ;

[0038] in, This represents the total number of data points within the sliding window; It is the wind speed forecast value at the i-th moment within the window; This is the arithmetic mean of all wind speed forecasts within the window. A higher fluctuation range indicates unstable environmental conditions, which may lead to drastic fluctuations in power generation. The second type is the trend indicator, which, in engineering terms, identifies the main direction and rate of change of meteorological factors in the short term. This trend indicator is obtained by performing linear regression analysis on the data series within the window and taking the slope of the fitted line. For calculating the trend indicator... ,have:

[0039] ;

[0040] in, This represents the time coordinate corresponding to the i-th data point within the window; This represents the weather forecast value at that moment. A trend indicator with a large absolute value represents a significant upward or downward trend, such as a wind power ramp-up event or a solar eclipse. Finally, the fluctuation amplitude indicators calculated for multiple key meteorological variables such as wind speed and irradiance are combined with the trend indicators to form a vector. This vector is the multidimensional environmental feature code. As a compact numerical summary, this multidimensional environmental feature code contains both information on environmental fluctuation risks and their evolution trend information. It is directly input into the dual-channel dynamic interval generator as the core basis for assessing environmental risks and adjusting the forecast interval width.

[0041] For example, we first obtain forecast data for the next 24 to 72 hours from a numerical weather prediction service provider. Suppose that at a specific time point, we obtain wind speed forecast data for a wind farm for the next hour, with a time resolution of 15 minutes. We analyze this 1-hour period (4 data points) as a sliding window. Assume the obtained wind speed forecast sequence is: 8.5 m / s at the current time +15 minutes, 8.8 m / s at the current time +30 minutes, 8.0 m / s at the current time +45 minutes, and 8.7 m / s at the current time +60 minutes. First, we calculate the wind speed fluctuation index within this window. (m / s), the result is rounded to three decimal places. Next, the wind speed trend index within this window is calculated. The time coordinates are normalized or encoded, for example, t1=1, t2=2, t3=3, t4=4, corresponding to a 15-minute interval. The weather forecast value is then... Calculate the trend indicator Similarly, for other key meteorological variables such as irradiance and cloud cover rate of change, corresponding fluctuation amplitude and trend indices are calculated within the same window. These indices are then combined into a multidimensional environmental feature code vector. By performing sliding window analysis on raw high-dimensional numerical weather prediction data and calculating fluctuation amplitude and trend indices, complex raw meteorological information can be effectively extracted into compact and physically meaningful multidimensional environmental feature codes, improving the efficiency and accuracy of the prediction model.

[0042] Based on the numerical weather forecast data, the preset base point prediction model is invoked to obtain the point prediction results and the base probability distribution;

[0043] Optionally, obtaining the point prediction results and the basic probability distribution includes:

[0044] The numerical weather forecast data is input into the base point prediction model, and meteorological and temporal correlation features are extracted and dynamically weighted to obtain a fused feature vector.

[0045] Based on the fused feature vector, power trend deduction is performed to obtain point prediction results;

[0046] The uncertainties in the power trend projection process are quantitatively estimated, and a basic probability distribution is formed through probability diffusion and confidence propagation mechanisms.

[0047] Specifically, the acquired numerical weather forecast data is first input into the baseline prediction model. Within this model, feature extraction is performed first. This process extracts two types of features in parallel. The first type is meteorological correlation features, which are feature vectors obtained by nonlinearly transforming and combining input meteorological data such as wind speed, irradiance, temperature, and humidity. These aim to capture the direct physical mapping relationship between meteorological conditions and power generation. The second type is temporal correlation features, which are feature vectors obtained by encoding the time information of the prediction time, such as hour, weekday, and month, using methods like one-hot encoding or periodic function encoding. These aim to capture time-dependent patterns such as the periodicity of power plant operation, maintenance plans, or grid dispatch instructions. Next, dynamic weight fusion is performed. The engineering meaning of this step is to intelligently adjust the importance of meteorological correlation features and temporal correlation features based on the current prediction context, generating a fused feature vector with higher information content. For example, in photovoltaic prediction, the weight of meteorological features during the day is much higher than that of temporal features, while the opposite is true at night. This fusion process can be implemented by a small attention network that dynamically generates weights. For calculating the fused feature vector at time t... ,have:

[0048] ;

[0049] in, and These are the extracted meteorological correlation feature vector and the time correlation feature vector, respectively. and These are meteorological and temporal weighting coefficients, generated by a small attention network based on the current input context, and their sum is 1. The relative importance of these two types of features is intelligently adjusted according to the prediction context; for example, in photovoltaic forecasting, the importance of time-related factors is adjusted based on the time of day. Larger, at night The power prediction is relatively large. After obtaining the fused feature vector, the system performs power trend extrapolation to generate point prediction results. This is typically accomplished by a deep learning model, such as a Long Short-Term Memory (LSTM) network, which takes the fused feature vector sequence as input and extrapolates the power value at future time points. To obtain a stable point prediction result, it is usually defined as the mean of multiple random forward propagation results. This randomness is introduced during the model inference phase using techniques such as Monte Carlo dropout. While extrapolating the point prediction results, uncertainty quantification estimation is performed in parallel, the core of which is to quantify the model's confidence in the current prediction. Using the Monte Carlo dropout technique, multiple Monte Carlo random forward propagations are performed on the same input fused feature vector, for example, 50 times, with random node dropouts, resulting in a set of slightly different power prediction values ​​containing that number of propagations. The point prediction result is the arithmetic mean of this set. Finally, using this set of prediction values, a basic probability distribution is formed through probability diffusion and confidence propagation mechanisms. In engineering, probability diffusion refers to using the point prediction result as the center of the distribution and using the dispersion of the prediction value set as a measure of uncertainty. Specifically, the standard deviation of this set is calculated, which directly reflects the volatility of the model output, i.e., the uncertainty of the prediction. The calculation of this standard deviation... ,have:

[0050] ;

[0051] in, The number of Monte Carlo random forward propagations; It is the predicted value obtained from the kth random forward propagation; The prediction result is then used as the center. A parameterized base probability distribution is then constructed based on this center and an uncertainty measure. Typically, this distribution is assumed to follow a normal distribution. For example, if a normal distribution is used, the base probability distribution can be represented as having a mean of [missing value]. The variance is The probability density function. This basic probability distribution provides the initial basis for subsequent probability correction and interval optimization.

[0052] For example, the acquired numerical weather forecast data is used as input and fed into a pre-defined baseline prediction model. Assume that in a certain prediction task, meteorological correlation feature vectors and time correlation feature vectors are extracted. For example, the meteorological correlation feature vector at a certain moment is [wind speed, irradiance, temperature] = A certain encoding representation, assuming the result is a vector of dimension 10, simplified here as A dimensionless vector. A time-related feature vector can be a representation of the predicted time's hour, day of the week, month, etc., after one-hot encoding or periodic function encoding, for example... This is a dimensionless vector. Next, dynamic weight fusion is performed. Assuming the current prediction context, such as daytime with ample sunlight, a small attention network generates meteorological and temporal weight coefficients of 0.7 and 0.3 respectively based on the input context. Then, the fused feature vector... Then, the fused feature vector is input into a deep learning model (such as LSTM) to extrapolate the power trend, and 50 random forward propagations are performed using Monte Carlo dropout. Assume these 50 propagations produce 50 slightly different watt-level power predictions. If, after 50 Monte Carlo random forward propagations, the simulated point prediction is 200.0 MW, the standard deviation of this set of predictions is 2.5 MW. Finally, a basic probability distribution is formed through probability diffusion and confidence propagation mechanisms. If this distribution follows a normal distribution, the basic probability distribution can be described as a normal distribution with a mean of 200.0 MW and a standard deviation of 2.5 MW. By transforming numerical weather prediction data into a fused feature vector and employing a deep learning model combined with Monte Carlo dropout, this method provides initial decision-making information with uncertainty for subsequent probability correction and prediction interval optimization, enhancing the comprehensiveness and reliability of the prediction results.

[0053] The multidimensional environmental feature code and the point prediction result are input into a dual-channel dynamic interval generator to generate a preliminary prediction interval.

[0054] Optionally, generating the preliminary prediction interval includes:

[0055] In the dual-channel dynamic interval generator, the difference analysis of the point prediction results at the same time is performed based on multiple sub-models in the basic point prediction model to generate a prediction confidence signal.

[0056] Environmental risk assessment is performed based on the multidimensional environmental feature code to generate environmental risk signals;

[0057] The predicted confidence signal and the environmental risk signal are fused to obtain a fusion result, and a dynamic interval is constructed based on the fusion result and the point prediction result to generate a preliminary prediction interval.

[0058] Specifically, the dual-channel dynamic interval generator is a parallel processing structure that receives point prediction results and multi-dimensional environmental feature codes as core inputs and outputs preliminary upper and lower limits for power prediction. In the first channel of the dual-channel dynamic interval generator, prediction confidence assessment is performed. This process utilizes multiple sub-models typically contained within the base point prediction model, such as 3 to 5 neural networks with different initialization parameters or structural variants within an ensemble learning framework. For the same future prediction time, all sub-models are driven to independently generate their point prediction results. The prediction confidence signal is defined as the degree of difference between the prediction results of these sub-models, typically quantified using standard deviation in engineering. The calculation of the prediction confidence signal... ,have:

[0059] ;

[0060] in, The number of sub-models; It is the point prediction result given by the i-th sub-model; This is the arithmetic mean of the prediction results from all sub-models. A smaller prediction confidence signal value indicates that the sub-models reach a consensus on the prediction results, indicating high model confidence; conversely, a larger value indicates significant internal disagreement within the model, indicating low confidence. In the second channel, environmental risk assessment is performed in parallel. This process receives a previously generated multidimensional environmental feature code as input. This multidimensional environmental feature code already includes quantitative indicators such as wind speed fluctuation amplitude and irradiance variation trend. Through a small feedforward neural network, this multidimensional vector is mapped into a single scalar signal, namely the environmental risk signal. Its engineering implication is to assess the stability of future meteorological conditions; unstable meteorological conditions, such as sudden changes in wind speed or rapid cloud movement, will lead to a higher environmental risk signal. Subsequently, the fusion and interval construction stage begins. The prediction confidence signals and environmental risk signals from the two channels are weighted and fused. Both are indicators measuring power fluctuation amplitude and have the same dimensions, resulting in a comprehensive fusion result. The calculation of the fusion result... ,have:

[0061] ;

[0062] in, and It is a dimensionless fusion weight coefficient, whose value can be between 0 and 1. It is obtained by training and optimization with historical data and is used to balance the relative importance of model uncertainty and environmental uncertainty. This serves as an environmental risk signal. Finally, a preliminary prediction interval is dynamically constructed based on the fusion result and the point prediction result. This interval is centered on the point prediction result, and its width is determined by the fusion result. For the preliminary prediction interval... ,have:

[0063] ;

[0064] in, It is a dimensionless confidence level coefficient; for example, for a 95% confidence level, the z-value is typically 1.96. This initial prediction interval can simultaneously reflect the model's confidence level and the level of challenge from the external environment.

[0065] For example, the previously obtained dual-channel dynamic interval generator receives point prediction results and multi-dimensional environmental feature codes as core inputs. If the point prediction result is 200.0 MW, a prediction confidence assessment is performed in the first channel of the dual-channel dynamic interval generator. Assuming the base point prediction model contains three sub-models whose point prediction results at the same time point are 199.5 MW, 201.0 MW, and 198.8 MW respectively, the arithmetic mean of all sub-model prediction results is approximately 199.77 MW. Calculation... (MW), the result is rounded to three decimal places. In the second channel, environmental risk assessment is performed in parallel. Assuming the previously generated multidimensional environmental feature code, such as wind speed fluctuations and irradiance trends, is mapped through a small feedforward neural network, it generates an environmental risk signal of 1.5MW. Subsequently, the prediction confidence signal and the environmental risk signal are weighted and fused to obtain a comprehensive fusion result. Assuming the fusion weight coefficients α=0.6 and β=0.4, obtained through optimization using historical data, the fusion result I is calculated. fused =0.6·1.124+0.4·1.5=1.2744 (MW). Finally, based on this fusion result and the point prediction result, a preliminary prediction interval is dynamically constructed. If the confidence level coefficient is 1.96, then PI pre =[200MW-1.96·1.2744MW,200MW+1.96·1.2744MW]=[197.5022MW,202.4978MW]. Through a dual-channel dynamic interval generator, the uncertainties of the model itself and the external environment were successfully quantified and integrated, providing a solid foundation for subsequent accuracy correction and practical deployment, and improving the adaptability and reliability of the prediction area.

[0066] Collect actual power data of new energy power generation and corresponding historical forecast data, calculate short-term error sequence, perform trend analysis on the short-term error sequence, and obtain short-term error trend characteristics;

[0067] Optionally, obtaining the short-term error trend characteristics includes:

[0068] The difference between the actual power and the historical predicted power is calculated point by point to form a short-term error sequence arranged in chronological order;

[0069] Local aggregation analysis within a sliding window is performed on the short-term error sequence to identify the direction and rate of change of the error in adjacent time periods;

[0070] Based on the direction and rate of change, a trend description vector reflecting the continuity and inertia of error evolution is constructed as a short-term error trend feature.

[0071] Specifically, the data acquisition and preprocessing process is initiated first, retrieving the most recent time period, such as the past 6 to 12 hours, of actual power generation data from new energy sources and corresponding historical forecast data from the historical database. The actual power data typically originates from the data acquisition and monitoring control system of the power plant, with a time resolution of 15 minutes, while the historical forecast data is previously generated archived data. The difference between the two is calculated point by point, representing the forecast error, thus forming a short-term error sequence arranged in chronological order. Next, a local aggregation analysis is performed within a sliding window on this short-term error sequence. This is implemented in engineering by using a fixed-length analysis window, for example, containing eight consecutive error data points from the past 1 to 2 hours, sliding point by point across the short-term error sequence. At each window position, aggregation calculations are performed to identify the direction and rate of error change within adjacent time periods. The rate of change is quantified by calculating the average difference between consecutive error values ​​within the window, reflecting the average rate of increase or decrease in error magnitude. The calculation of the error change rate... ,have:

[0072] ;

[0073] in, and These are the last and first error values ​​within the current sliding window, respectively. This refers to the number of data points within the window. The sign of the error change rate directly reflects the direction of error change, i.e., whether the error is increasing or decreasing; its absolute value represents the drasticness of the change, i.e., the rate of change. Finally, based on the above analysis results, a trend description vector that reflects the continuity and inertia of error evolution is constructed; this vector is the short-term error trend characteristic. This vector contains at least two core components: one is the average error within the current window. The first represents the average systematic bias level of the current prediction; the second is the calculated error change rate. Therefore, the short-term error trend characteristics generated at time t can be constructed as follows: This two-dimensional vector concisely characterizes the static offset and dynamic evolution trend of recent errors. A consistently positive short-term error trend with a positive error rate indicates a persistent and escalating underestimation tendency in the model. This inertial information will be used to guide subsequent probability distribution corrections. Figure 2As shown, subplot (a) shows the trend of actual power and historical predicted power over time, intuitively showing the difference between the predicted and actual values; subplot (b) calculates and plots the short-term error sequence, that is, the point-by-point difference between actual power and historical predicted power. A sample sliding window is marked with a gray area in the figure. The trend line and its slope within the window quantify the direction and rate of change of the error, which together constitute the short-term error trend characteristics.

[0074] For example, the actual power generation data of new energy sources over the most recent 12 hours (e.g., from the historical prediction time minus 12 hours to the current time) and the corresponding historical prediction data are retrieved from a historical database. The time resolution is 15 minutes, resulting in a total of 48 data points. Assuming a sliding window contains, for example, four data points from the most recent hour, the actual power and historical prediction power are obtained, and the difference between the actual power and the historical prediction power is calculated point by point; this is the prediction error. Next, a local aggregation analysis is performed on this short-term error sequence within the sliding window. The number of data points in this window is four. First, the average error within the current window is calculated to be 1.5. Considering that the data points within the window are evenly spaced, the first error value within the window is set to -2. The last error value in the window is 8. Then calculate the rate of change of error. ( The result is rounded to two significant figures, where step represents a 15-minute time step. Finally, based on the above analysis, a short-term error trend characteristic is constructed. By structurally collecting, serializing, and analyzing recent historical prediction errors through sliding window analysis, short-term error trend characteristics that reflect the systematic deviation level and dynamic evolution trend of the error can be effectively extracted, enabling the overall prediction system to quickly adapt to and correct potential prediction biases.

[0075] The basic probability distribution is adjusted using the short-term error trend characteristics to generate a corrected probability distribution;

[0076] Optionally, the generation of the corrected probability distribution includes:

[0077] Identify the co-directionality of continuous prediction errors in the short-term error sequence and generate systematic bias direction and intensity characteristics;

[0078] Based on the characteristics of the direction and intensity of the systematic deviation, determine the distribution shift and distribution shape adjustment.

[0079] The basic probability distribution is adjusted using the distribution shift and the distribution shape adjustment to generate a corrected probability distribution.

[0080] Specifically, when performing this task, the input short-term error sequence is first analyzed to identify the unidirectionality of continuous prediction errors, i.e., whether the errors consistently remain positive or negative. In engineering, this is achieved by checking the sign consistency of the nearest points in the short-term error sequence, for example, 5 to 10 points. If most error points have the same sign, a systematic bias is considered to exist. Based on this analysis, the direction and intensity characteristics of the systematic bias are generated. The direction characteristic is determined by the sign of the average error in the short-term error sequence, while the intensity characteristic is quantified by its absolute value. Next, based on the aforementioned systematic bias direction and intensity characteristics, two core parameters for adjusting the base probability distribution are determined: the distribution shift and the distribution shape adjustment. The distribution shift is directly used to compensate for the mean bias of the prediction; its value is directly related to the average error in the short-term error trend characteristics, but it is attenuated by a smoothing coefficient to prevent overcorrection caused by a single drastic error. The calculation of the distribution shift... ,have:

[0081] ;

[0082] in, It is a dimensionless smoothing coefficient, typically ranging from 0.1 to 0.5, and is obtained through optimization using historical data; E represents the number of error samples. i Let be the prediction error at time i. The distribution shape adjustment is used to adjust the dispersion of the probability distribution, i.e., the magnitude of uncertainty. It is mainly driven by the rate of error change in the short-term error trend characteristics. A continuously increasing error trend means that future uncertainty is increasing, requiring a wider distribution. The distribution shape adjustment can be a multiplicative adjustment factor of the standard deviation. For calculating the distribution shape adjustment... ,have:

[0083] ;

[0084] in, It is a dimensionless morphological adjustment sensitivity coefficient; It is the rated installed capacity of the power station, used for... Normalization is performed to make it a dimensionless quantity. Finally, the system uses these two parameters to adjust the basic probability distribution. Assume the basic probability distribution is a distribution with the point prediction results as the mean and the model uncertainty as the factor. The standard deviation follows a normal distribution. The adjustment process consists of two steps. First, the distribution shift is applied to the mean, resulting in a corrected mean that is the sum of the predicted point and the distribution shift. Second, the distribution shape adjustment is applied to the standard deviation, resulting in a corrected standard deviation that is the sum of the distribution shape adjustment and the standard deviation. The product of these factors. Ultimately, the resulting corrected probability distribution is a new normal distribution that simultaneously compensates for mean bias and dynamically adjusts the estimation of uncertainty. For example... Figure 3 As shown in the figure, the basic probability distribution and the corrected probability distribution are superimposed. The comparison of the two probability density curves clearly reflects the distribution changes caused by the short-term error trend characteristics. Compared with the basic distribution, the corrected distribution has undergone an overall positional shift and an adjustment in its dispersion, i.e., the standard deviation has changed.

[0085] For example, the initial base probability distribution is dynamically adjusted using short-term error trend features extracted from recent historical deviations. The previously obtained base probability distribution is a normal distribution with a mean of 200.0 MW and a standard deviation of 2.5 MW. Assume that by analyzing the past 10 error points, systematic deviations are identified and the smoothed average error is calculated, which is used as the distribution shift. Assume a smoothing coefficient of 0.2, and the sum of the recent 10 error points is 75 MW. First, the distribution shift is calculated. (MW). Next, the distribution pattern adjustment amount is determined based on the error change rate. If the rated installed capacity of the wind power station is 200MW and the pattern adjustment sensitivity coefficient is 0.5step, then the distribution pattern adjustment amount is calculated. The results are rounded to two decimal places. Finally, these two parameters are used to adjust the basic probability distribution, resulting in a new normal probability distribution. By combining the direction and intensity characteristics of systematic deviations and utilizing the two core parameters of distribution shift and distribution shape adjustment, the basic probability distribution can be accurately and dynamically corrected, improving the accuracy and real-time adaptability of probability prediction and providing more reliable probability information for subsequent optimization of the prediction interval.

[0086] Optionally, adjusting the basic probability distribution using the distribution shift and the distribution shape adjustment includes:

[0087] Based on the distribution translation amount, the overall position of the basic probability distribution is shifted to compensate for the systematic deviation;

[0088] After completing the overall positional offset, the dispersion and skewness characteristics of the basic probability distribution are reshaped according to the distribution shape adjustment amount;

[0089] An adaptive smoothing mechanism is introduced during the adjustment process to constrain overcorrection caused by drastic local fluctuations, resulting in a corrected probability distribution.

[0090] Specifically, the first step involves performing a global position shift operation based on the distribution shift. This step aims to directly compensate for systematic biases in the model. The previously calculated distribution shift is applied directly to the expected value or center position of the underlying probability distribution. If the underlying probability distribution is parameterized by its mean and standard deviation, this step will result in a better understanding of the updated distribution's mean. ,have:

[0091] ;

[0092] in, This is the mean of the basic probability distribution, i.e., the point prediction result. This addition operation is valid in the physical dimension because all three are units of power. After completing the overall position shift, the dispersion and skewness of the probability distribution are reshaped according to the distribution shape adjustment. This operation mainly affects the width of the distribution, i.e., the range of uncertainty in the prediction. The corrected standard deviation is obtained by multiplying the basic standard deviation by the dimensionless distribution shape adjustment. The distribution shape adjustment reflects the additional uncertainty brought about by the recent error evolution trend. When the error trend intensifies, the distribution shape adjustment is greater than 1, widening the distribution; conversely, it narrows the distribution. In more advanced implementations, if a skew-supporting distribution model such as a skewed normal distribution is used, the distribution shape adjustment can also be a vector, used to simultaneously adjust the skewness parameter of the distribution to reflect the asymmetry of the error. Throughout the adjustment process, an adaptive smoothing mechanism is introduced at the core, the engineering purpose of which is to constrain overcorrection caused by local drastic fluctuations, such as a single anomalous error value. This mechanism is achieved by dynamically adjusting the update rate of the distribution shift and the distribution shape adjustment. For example, the smoothing coefficient used in calculating the distribution shift is no longer a fixed value, but is adaptively adjusted based on the variance of the recent error sequence. When the variance of the error sequence is large, indicating drastic and unstable fluctuations, the smoothing coefficient is automatically reduced to decrease the impact of the current error on the overall distribution shift; when the error sequence is stable and consistently in the same direction, the smoothing coefficient is appropriately increased to accelerate compensation for systematic deviations. This adaptive damping ensures the smoothness and robustness of the correction process, and the resulting corrected probability distribution more stably and accurately reflects future power fluctuation characteristics.

[0093] For example, the point prediction results calculated above are used as the mean and standard deviation of the basic probability distribution. The mean distribution is 2.5MW, the distribution shift is 1.5MW, and the distribution shape adjustment is 1.01. First, the system performs an overall position offset operation based on the distribution shift. The updated distribution mean... After completing the overall positional shift, the dispersion of the probability distribution is reshaped based on the adjustment amount of the distribution pattern. The corrected standard deviation is... An adaptive smoothing mechanism is introduced throughout the adjustment process. For example, the smoothing coefficient when calculating the distribution shift is no longer a fixed value. If the variance of the recent error sequence is small, indicating that the error is stable and consistently in the same direction, the smoothing coefficient will be automatically increased, for example, from 0.2 to 0.3, to accelerate compensation for systematic biases. If the variance of the error sequence is large, indicating drastic and unstable fluctuations, the smoothing coefficient will be automatically decreased to 0.1 to reduce the impact of the current error on the overall distribution shift and prevent overcorrection. In practice, the smoothing coefficient value can be dynamically determined using a lookup table based on historical error variance or a neural network. By precisely applying the distribution shift and distribution shape adjustment, supplemented by the adaptive smoothing mechanism, the systematic bias of the model is effectively compensated, and the prediction uncertainty is flexibly adjusted according to the dynamic evolution trend of the error. Simultaneously, the adaptive mechanism avoids overcorrection caused by local fluctuations.

[0094] The probability verification signal is extracted from the corrected probability distribution and fed back to the dual-channel dynamic interval generator. The preliminary prediction interval is optimized and adjusted based on the probability verification signal to generate and output the optimized prediction interval.

[0095] Optionally, generating and outputting the optimized prediction interval includes:

[0096] The probability verification signal is extracted from the corrected probability distribution, and the probability verification signal is analyzed to obtain the interval translation guidance and the interval width adjustment guidance.

[0097] The interval translation guide and the interval width adjustment guide are then fused with the fusion result to obtain a secondary fusion result.

[0098] Based on the secondary fusion results, the position and width of the initial prediction interval are adjusted to generate the optimized prediction interval.

[0099] Specifically, the probability verification signal is first extracted from the generated corrected probability distribution. The extraction process quantifies the difference between the base probability distribution and the corrected probability distribution. Two core components are parsed in parallel: an interval shift guide and an interval width adjustment guide. The interval shift guide is numerically equal to the distribution shift previously calculated for correcting the probability distribution; it directly quantifies the offset required to the center position of the interval due to systematic bias, and its unit is consistent with the power unit. The interval width adjustment guide is obtained by calculating the ratio of the standard deviation of the corrected probability distribution to the standard deviation of the base probability distribution; it is a dimensionless scaling factor representing the scaling ratio required for prediction uncertainty. Next, these two guides are fused a second time with the preliminary fusion result calculated in the dual-channel dynamic interval generator. This second fusion is an application process, not a simple mathematical superposition. In engineering implementation, the interval shift guide and the interval width adjustment guide are used as the final adjustment operators, applied to the position and width parameters of the preliminary prediction interval, respectively. The output of this process is defined as the secondary fusion result, which contains an updated interval center point and an updated interval half-width. The calculation of the updated interval center point... and the updated interval half-width ,have:

[0100] ;

[0101] in, This serves as a guide for interval translation; The system adjusts the interval width using guidance parameters. Finally, based on this secondary fusion result—the updated interval center point and interval half-width—the system generates and outputs the optimized prediction interval. Its construction method is similar to that of the initial interval generation, but it uses parameters adjusted through correction feedback. The calculation of the optimized prediction interval... ,have:

[0102] ;

[0103] The optimized prediction range of this output has its center position compensated for the short-term prediction mean deviation, and its width is the final uncertainty estimate obtained after comprehensively considering model uncertainty, environmental risk and error dynamic evolution trend, thus achieving comprehensive optimization of the new energy power prediction range.

[0104] For example, a probability verification signal is extracted from the corrected probability distribution and fed back to the dual-channel dynamic interval generator. If the mean of the base probability distribution is 200MW and the standard deviation is 2.5MW, and the mean of the corrected probability distribution is 201.5MW and the corrected standard deviation is 2.525MW, the probability verification signal is first extracted from the corrected probability distribution to obtain the interval translation guidance and the interval width adjustment guidance. The interval translation guidance is then calculated. (MW). Calculate the guidance amount for interval width adjustment. This is a dimensionless scaling factor, representing the proportion by which the prediction uncertainty needs to be scaled. These guidance values ​​are then fused a second time with the fusion result previously calculated in the dual-channel dynamic interval generator. If the initial fusion result is 1.2744MW, the updated interval center point P is calculated. final =200+1.5=201.5 (MW), calculate the updated interval half-width I. final =1.2744·1.01≈1.287 (MW), the result is rounded to three decimal places. Based on this secondary fusion result, the optimized prediction interval is generated and output. The confidence level coefficient is set to 1.96, then Piopt=[201.5-1.96·1.287,201.5+1.96·1.287]=[198.97748,204.02252] (MW). By extracting precise interval shift guidance and interval width adjustment guidance from the corrected probability distribution and feeding them back to the interval generation process, the system performs a final fine correction on the initial prediction interval, improving the accuracy, adaptability and robustness of the new energy power prediction interval.

[0105] Optionally, extracting the probability verification signal from the corrected probability distribution includes:

[0106] Calculate the overall offset of the corrected probability distribution relative to the basic probability distribution;

[0107] Based on the overall offset and the characteristics of the direction and intensity of the systematic deviation, a probability verification signal is generated and sent to the dual-channel dynamic range generator.

[0108] Specifically, when performing this task, the overall offset of the corrected probability distribution relative to the base probability distribution is first calculated. This operation is achieved by accessing and comparing the core positional parameters of the two probability distributions, namely their means. The overall offset is numerically equal to the arithmetic difference between the mean of the corrected probability distribution and the mean of the base probability distribution. This calculation directly quantifies the amount of translation compensation to the distribution center due to systematic bias, and its unit is power, with a clear physical meaning. Subsequently, based on this overall offset and the previously obtained direction and intensity characteristics of the systematic bias, a composite probability check signal is generated. This signal is not a single numerical value, but a structured information packet containing multiple instruction components, typically implemented as a low-dimensional vector. The primary component of this probability check signal is the just-calculated overall offset, which directly carries the instruction to translate the position of the interval center. To make the signal complete, the system also encapsulates an instruction component representing width adjustment, which originates from the influence of the systematic bias characteristics on uncertainty. For example, this component could be the ratio of the standard deviation of the corrected probability distribution to the standard deviation of the base probability distribution, which is a dimensionless scaling factor. Finally, this probability verification signal, which contains translation and scaling instructions, is sent to the dual-channel dynamic range generator via an internal message queue or function call, serving as the direct input for its range optimization adjustment.

[0109] For example, a probability check signal is extracted from the corrected probability distribution to quantify and encapsulate the result of the probability distribution adjustment, forming a transferable correction instruction. Assume the mean of the base probability distribution is 200.0 MW and the standard deviation is 2.5 MW, while the mean of the corrected probability distribution is 201.5 MW and the standard deviation is 2.525 MW. First, the overall offset of the corrected probability distribution relative to the base probability distribution is calculated. This offset is obtained by subtracting the mean of the base probability distribution from the mean of the corrected probability distribution, for example, 1.5 MW. This value directly quantifies the translation compensation required for the distribution center due to systematic deviation; its unit is consistent with the power unit, and its physical meaning is clear. Subsequently, based on this overall offset and combined with the pre-obtained direction and intensity characteristics of the systematic deviation, a composite probability check signal is generated. The primary component of this signal is the calculated overall translation amount of 1.5 MW, which directly carries the instruction to shift the position of the interval center. To ensure signal integrity, the signal also encapsulates a width adjustment instruction component, obtained by the ratio of the standard deviation of the corrected probability distribution to the standard deviation of the base probability distribution, for example, approximately 1.01. Finally, this probability verification signal, containing translation and scaling instructions, is encapsulated and passed through an internal message queue to the dual-channel dynamic interval generator as direct input for its interval optimization adjustments. By accurately calculating and encapsulating the overall offset of the corrected probability distribution relative to the base probability distribution and the width adjustment ratio, seamless information transfer between probability distribution correction and interval generation is achieved, improving the accuracy and effectiveness of the final predicted interval.

[0110] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides an interval generation and probability correction system for new energy power prediction, the system comprising:

[0111] The forecast feature generation module is used to acquire numerical weather forecast data and perform dynamic stability analysis based on the numerical weather forecast data to generate multi-dimensional environmental feature codes.

[0112] The point prediction distribution module is used to call a preset basic point prediction model to obtain point prediction results and basic probability distribution.

[0113] The interval preliminary generation module is used to input the multidimensional environmental feature code and the point prediction result into the dual-channel dynamic interval generator to generate a preliminary prediction interval.

[0114] The error trend analysis module is used to collect actual power data of new energy power generation and corresponding historical forecast data, calculate short-term error sequence, perform trend analysis on the short-term error sequence, and obtain short-term error trend characteristics.

[0115] The probability distribution correction module is used to adjust the basic probability distribution using the short-term error trend characteristics to generate a corrected probability distribution.

[0116] The interval optimization output module is used to extract the probability verification signal from the corrected probability distribution and feed it back to the dual-channel dynamic interval generator. Based on the probability verification signal, the module optimizes and adjusts the preliminary prediction interval, and generates and outputs the optimized prediction interval.

[0117] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0118] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for interval generation and probability correction of new energy power prediction, characterized in that: obtaining numerical weather prediction data, and performing dynamic stability analysis based on the numerical weather prediction data to generate a multi-dimensional environmental feature code; calling a preset basic point prediction model based on the numerical weather prediction data to obtain a point prediction result and a basic probability distribution; inputting the multi-dimensional environmental feature code and the point prediction result into a double-channel dynamic interval generator to generate a preliminary prediction interval; collecting actual power generation data and corresponding historical prediction data of new energy, calculating a short-term error sequence, performing trend analysis on the short-term error sequence to obtain a short-term error trend feature; adjusting the basic probability distribution using the short-term error trend feature to generate a corrected probability distribution; extracting a probability check signal from the corrected probability distribution and feeding it back to the double-channel dynamic interval generator, and optimizing and adjusting the preliminary prediction interval according to the probability check signal to generate and output an optimized prediction interval. The generation of the multi-dimensional environmental feature code comprises: obtaining numerical weather prediction data including wind speed, irradiance and cloud cover change rate; calculating fluctuation amplitude indicators and change trend indicators in the future period based on the numerical weather prediction data; and analyzing and calculating the fluctuation amplitude indicators and the change trend indicators to generate a multi-dimensional environmental feature code. The obtaining of the point prediction result and the basic probability distribution comprises: inputting the numerical weather prediction data into the basic point prediction model, extracting meteorological correlation features and time correlation features and performing dynamic weight fusion to obtain a fusion feature vector; performing power trend deduction according to the fusion feature vector to obtain a point prediction result; and quantitatively estimating the uncertainty in the power trend deduction process to form a basic probability distribution through probability diffusion and confidence propagation mechanism. The generation of the preliminary prediction interval comprises: in the double-channel dynamic interval generator, performing difference analysis on the point prediction result at the same time based on a plurality of sub-models in the basic point prediction model to generate a prediction confidence signal; performing environmental risk assessment based on the multi-dimensional environmental feature code to generate an environmental risk signal; fusing the prediction confidence signal and the environmental risk signal to obtain a fusion result, and performing dynamic interval construction based on the fusion result and the point prediction result to generate a preliminary prediction interval. The obtaining of the short-term error trend feature comprises: point-by-point calculating the difference between the actual power and the historical predicted power to form a short-term error sequence arranged in chronological order; performing local aggregation analysis on the short-term error sequence within a sliding window to identify the change direction and change rate of the error in adjacent periods; and based on the change direction and change rate, constructing a trend description vector reflecting the continuity and inertia of error evolution as a short-term error trend feature. The generation of the corrected probability distribution comprises: identifying the same direction of continuous prediction errors in the short-term error sequence to generate a systematic deviation direction and intensity feature; determining a distribution shift amount and a distribution form adjustment amount according to the systematic deviation direction and intensity feature; ​ 2. The interval generation and probability correction method for new energy power prediction according to claim 1, characterized in that, ​ ​ ​ ​ 3. The interval generation and probability correction method for new energy power prediction according to claim 1, characterized in that, ​ ​ ​ ​ 4. The interval generation and probability correction method for new energy power prediction according to claim 1, characterized in that, ​ ​ ​ ​ 5. The interval generation and probability correction method for new energy power prediction according to claim 1, characterized in that, ​ ​ ​ ​ 6. The interval generation and probability correction method for new energy power prediction according to claim 4, characterized in that, ​ ​ ​ The base probability distribution is adjusted by using the distribution translation amount and the distribution shape adjustment amount to generate a corrected probability distribution.

7. The interval generation and probability correction method for new energy power prediction according to claim 6, characterized in that, The adjusting the base probability distribution by using the distribution translation amount and the distribution shape adjustment amount comprises: compensating for the systematic deviation by performing overall position offset of the base probability distribution based on the distribution translation amount; reshaping the dispersion degree and skewness characteristics of the base probability distribution according to the distribution shape adjustment amount after the overall position offset is completed; introducing an adaptive smoothing mechanism in the adjustment process to constrain over-correction caused by local sharp fluctuations, and obtaining the corrected probability distribution.

8. The interval generation and probability correction method for new energy power prediction according to claim 6, characterized in that, The generating and outputting the optimized prediction interval comprises: extracting a probability check signal from the corrected probability distribution, analyzing the probability check signal to obtain an interval translation guide and an interval width adjustment guide; performing secondary fusion of the interval translation guide and the interval width adjustment guide with the fusion result to obtain a secondary fusion result; adjusting the position and width of the preliminary prediction interval based on the secondary fusion result to generate an optimized prediction interval.

9. The interval generation and probability correction method for new energy power prediction according to claim 8, characterized in that, The extracting a probability check signal from the corrected probability distribution comprises: calculating the overall offset of the corrected probability distribution relative to the base probability distribution; generating a probability check signal according to the overall offset and the direction and intensity characteristics of the systematic deviation, and sending the probability check signal to the dual-channel dynamic interval generator.

10. A system for interval generation and probability correction of new energy power prediction, applied to the method for interval generation and probability correction of new energy power prediction according to any one of claims 1-9, characterized in that, The system comprises: a prediction feature generation module configured to obtain numerical weather prediction data and perform dynamic stability analysis based on the numerical weather prediction data to generate a multi-dimensional environmental feature code; a point prediction distribution module configured to call a preset base point prediction model to obtain a point prediction result and a base probability distribution; an interval preliminary generation module configured to input the multi-dimensional environmental feature code and the point prediction result into a dual-channel dynamic interval generator to generate a preliminary prediction interval; an error trend analysis module configured to collect new energy power generation actual power data and corresponding historical prediction data, calculate a short-term error sequence, and perform trend analysis on the short-term error sequence to obtain short-term error trend characteristics; a probability distribution correction module configured to adjust the base probability distribution by using the short-term error trend characteristics to generate a corrected probability distribution; an interval optimization output module configured to extract a probability check signal from the corrected probability distribution and feed back to the dual-channel dynamic interval generator, and optimize and adjust the preliminary prediction interval according to the probability check signal to generate and output an optimized prediction interval.

Images