Sub-daylight photovoltaic generating capacity optimization prediction method for electricity market transaction

Abstract

The invention, which belongs to the technical field of new energy power generation prediction, discloses a power market transaction-oriented secondary sunlight photovoltaic power generation optimization prediction method comprising the steps of obtaining historical data to construct a historical situation vector, performing reverse deduction and optimization on a historical declaration curve, and establishing a situation-strategy mapping library; generating a situation vector for the prediction day, and retrieving a candidate strategy set from a mapping library through similarity matching; outputting an economic benefit and risk multi-dimensional diagnosis report of the candidate strategy, and carrying out replacement and recombination on the strategy fragment based on contradiction diagnosis to generate a derivative strategy; and screening an optimal target declaration strategy for the candidate and derivative strategy scores, and updating the mapping library and the strategy generation model according to actual feedback after the trading day. According to the method, a data-driven situation matching and strategy optimization mechanism is adopted, and multi-source information fusion and dynamic strategy recombination are combined, so that declaration income optimization and risk control of the photovoltaic power station in the power market can be realized, and the practicability and economical efficiency of a prediction result are improved.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation forecasting technology, and in particular to an optimized forecasting method for next-day photovoltaic power generation for electricity market transactions. Background Technology

[0002] Photovoltaic power generation, as a clean energy source, plays an increasingly important role in the global energy structure transformation. With the deepening of electricity market reforms, photovoltaic power plants and other new energy power generation entities need to participate as independent market participants in day-ahead electricity market bidding. Under this model, power plants need to submit their planned power generation for each time slot of the following day—the declared power curve—to the market operator one day in advance. This declared power curve not only determines the power plant's basic revenue, but the deviation between it and the actual power generation will also generate additional economic penalties or rewards under market rules, directly affecting the power plant's final profitability.

[0003] In related technologies, Chinese invention patent CN119228427A discloses a method, apparatus, and equipment for generating day-ahead declaration optimization strategies for the electricity spot market. The method includes: preprocessing acquired operational boundary data and historical transaction data of the electricity spot market in a target area to obtain processed data; optimizing a pre-trained initial electricity price prediction model based on the processed data and historical electricity price prediction data to obtain an electricity price prediction model; generating at least one day-ahead declaration optimization strategy based on acquired user electricity demand information and day-ahead electricity price prediction data obtained from the electricity price prediction model; and selecting a target day-ahead declaration optimization strategy from the at least one day-ahead declaration optimization strategy based on the analysis results obtained from analyzing each of the at least one day-ahead declaration optimization strategies.

[0004] However, the aforementioned existing technical solutions have the following technical shortcomings. The core of this solution lies in optimizing the electricity price forecasting model and generating a reporting strategy based on the predicted electricity price. However, for photovoltaic power plants, the biggest challenge comes from the prediction deviation of power generation caused by the uncertainty of meteorological conditions. Existing solutions fail to deeply integrate key information such as the degree of divergence in meteorological forecasts and historical deviation risks, and fail to construct an effective risk quantification and control mechanism specifically for the volatile characteristics of photovoltaic power generation. The strategies generated by this solution are static and lack an effective diagnostic and optimization process. It cannot perform multi-dimensional and refined diagnosis of the economics and risks of the initially generated candidate strategies, nor can it perform fragmented and targeted replacement and reorganization of strategies based on the diagnostic results. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method for optimizing and predicting next-day photovoltaic power generation for electricity market transactions. It employs a data-driven scenario matching and strategy optimization mechanism, combined with multi-source information fusion and dynamic strategy reorganization, which enables the optimization of photovoltaic power plant declaration revenue and risk control in the electricity market, thereby improving the practicality and economic efficiency of the prediction results.

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

[0007] A method for optimizing and predicting photovoltaic power generation for the next day in the electricity market includes: acquiring historical data to construct a historical scenario vector; reverse-engineering and optimizing historical application curves and establishing a scenario-strategy mapping library; generating scenario vectors for the prediction day and retrieving a set of candidate strategies from the mapping library through similarity matching; outputting a multi-dimensional diagnostic report on the economic benefits and risks of the candidate strategies; replacing and recombining strategy fragments based on contradiction diagnosis to generate derivative strategies; scoring the candidate and derivative strategies to select the optimal target application strategy; and updating the mapping library and strategy generation model after the trading day based on actual feedback.

[0008] Optionally, the calculation of the historical scenario vector includes: extracting numerical weather forecasts from different sources from the multi-source meteorological data, calculating their differences, and generating a meteorological forecast divergence index; extracting time-of-use electricity price sequences from the electricity market data, calculating their intraday fluctuation range, and generating an electricity price volatility intensity index; analyzing historical daily deviations similar to the historical scenario vector from the power plant operation data, and combining them with the market assessment rules of the same period to calculate a historical deviation risk index; and combining the meteorological forecast divergence index, the electricity price volatility intensity index, and the historical deviation risk index into a historical scenario vector.

[0009] Optionally, the process of retrieving the candidate strategy set includes: calculating the weighted spatial distance between the predicted daily scenario vector and each historical scenario vector in the scenario-strategy mapping library, and converting the distance into a similarity score; sorting the historical scenario vectors in descending order according to the similarity score, and selecting several of the top-ranked historical scenario vectors as a set of similar historical scenario vectors; and extracting the optimized historical application curves that are uniquely associated with each vector in the set of similar historical scenario vectors from the scenario-strategy mapping library to form a candidate strategy set.

[0010] Optionally, the output, which includes a multi-dimensional contradiction diagnosis report on expected economic benefits and risk indicators, includes: performing multiple simulations on each candidate application curve to obtain a set of deviation power sequences; dividing the deviation power sequences into a first deviation distribution for high-price periods and a second deviation distribution for low-price periods based on the time-of-use electricity price on the prediction date; calculating the first risk cost of insufficient power generation during high-yield periods based on the first deviation distribution and preset negative deviation assessment rules; calculating the second resource waste cost of excessive power generation during low-yield periods based on the second deviation distribution and the market on-grid electricity price; and integrating the first risk cost and the second resource waste cost to output a multi-dimensional contradiction diagnosis report on expected economic benefits and risk indicators.

[0011] Optionally, generating the derivative strategy set includes: determining whether the first risk cost in the multi-dimensional contradiction diagnosis report exceeds a preset first risk threshold; if so, selecting a curve segment with a more conservative forecast value during the corresponding high electricity price period from the candidate strategy set, replacing the original curve segment, and generating a first type of derivative strategy; determining whether the second resource waste cost in the multi-dimensional contradiction diagnosis report exceeds a preset second opportunity cost threshold; if so, selecting a curve segment with a more optimistic forecast value during the corresponding low electricity price period from the candidate strategy set, replacing the original curve segment, and generating a second type of derivative strategy; and summarizing all the first type of derivative strategies and the second type of derivative strategies to form a derivative strategy set.

[0012] Optionally, updating the scenario-strategy mapping library and the strategy generation model for generating the target application strategy includes: calculating the actual comprehensive return of the target application strategy on the forecast date based on the actual market feedback data; calculating the optimized ex-post strategy for the day based on the complete market information and actual power generation data on the forecast date; associating the scenario vector of the forecast date with the optimized ex-post strategy to form a new mapping pair and storing it in the scenario-strategy mapping library; using the new mapping pair as training samples to incrementally train the strategy generation model and optimize its generation of the target application strategy based on the scenario vector.

[0013] Optionally, obtaining the corresponding multi-source meteorological data and electricity market data, and calculating the predicted daily scenario vector includes: performing feature extraction and standardization processing on the multi-source meteorological data and electricity market data to obtain a standardized feature sequence; performing principal component analysis and key indicator fusion on the standardized feature sequence to calculate and generate the predicted daily scenario vector.

[0014] Optionally, combining the meteorological forecast divergence index, the electricity price volatility index, and the historical deviation risk index into a historical scenario vector includes: normalizing the meteorological forecast divergence index, the electricity price volatility index, and the historical deviation risk index to obtain normalized indices; and performing a weighted linear combination of the normalized indices to generate a historical scenario vector.

[0015] Optionally, before constructing the context-policy mapping library, the method further includes: acquiring a large number of historical context vector samples and corresponding high-quality declaration curve samples to form a training set; defining an initial neural network model architecture containing an encoder and a decoder to learn the mapping relationship from context vectors to declaration curves; and using the training set to perform supervised training on the initial neural network model until the model performance converges, thereby obtaining a policy generation model that can generate policy genealogies based on input context vectors.

[0016] Based on the same inventive concept, this invention also provides a next-day photovoltaic power generation optimization prediction system for electricity market transactions. The system includes: a data acquisition and processing module, used to acquire multi-source meteorological data, electricity market data, and power plant operation data corresponding to historical trading days, and calculate and generate historical scenario vectors based on the acquired data; a scenario strategy library construction module, used to reverse-engineer optimized historical reporting curves based on the actual operating results of the historical trading days, and associate the historical scenario vectors with the optimized historical reporting curves to construct a scenario-strategy mapping library; a scenario vector calculation module, used to acquire the corresponding multi-source meteorological data and electricity market data for the prediction day, and calculate and generate a prediction day scenario vector; and a similar scenario matching and retrieval module, used to perform similarity matching between the prediction day scenario vector and the historical scenario vectors in the scenario-strategy mapping library to retrieve candidate scenarios. The system comprises the following modules: a strategy selection module; a multi-dimensional contradiction diagnosis module, used to output a multi-dimensional contradiction diagnosis report containing expected economic benefits and risk indicators based on each candidate application curve in the candidate strategy set; a strategy derivation and optimization module, used to replace and reorganize specific time segments of each candidate application curve according to the strategy contradictions revealed in the multi-dimensional contradiction diagnosis report, generating a derivative strategy set; a strategy evaluation and decision-making module, used to score each candidate application curve in the candidate strategy set and the derivative strategy set, and select the candidate application curve with the higher score as the target application strategy; and a feedback learning and model update module, used to obtain actual market feedback data containing actual settlement results after the end of the trading day, and update the scenario-strategy mapping library and the strategy generation model used to generate the target application strategy based on the actual market feedback data.

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

[0018] This invention constructs a mapping relationship between context and strategy, and combines similarity matching retrieval historical experience to provide a candidate application scheme that has been tested in practice for the prediction date. This avoids the huge uncertainty caused by simply relying on single-point physical or statistical predictions, and improves the initial quality and rationality of the application strategy.

[0019] This invention proposes a multi-dimensional contradiction diagnosis and strategy reorganization mechanism, which can perform in-depth economic analysis of candidate application curves, quantify their potential risk costs and opportunity costs in different electricity price periods, and perform targeted strategy fragment replacement and optimization based on the diagnosis results.

[0020] 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

[0021] 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.

[0022] Figure 1 This is a flowchart illustrating an embodiment of the present invention of a method for optimizing and predicting the next day's photovoltaic power generation for electricity market transactions.

[0023] Figure 2 This is a context vector space distribution and similarity matching graph according to an embodiment of the present invention.

[0024] Figure 3 This is a scenario similarity distribution diagram according to an embodiment of the present invention.

[0025] Figure 4 This is a schematic diagram of the structure of a next-day photovoltaic power generation optimization prediction system for electricity market transactions according to an embodiment of the present invention. Detailed Implementation

[0026] 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.

[0027] Reference Figure 1 One embodiment of the present invention proposes an optimized forecasting method for next-day photovoltaic power generation for electricity market transactions. It adopts a data-driven scenario matching and strategy optimization mechanism, combined with multi-source information fusion and dynamic strategy reorganization, which can optimize the declared revenue and control the risks of photovoltaic power plants in the electricity market, and improve the practicality and economy of the forecasting results.

[0028] The method described in this embodiment specifically includes:

[0029] S1. Obtain multi-source meteorological data, power market data, and power plant operation data corresponding to historical trading days, and calculate and generate historical scenario vectors based on the obtained data;

[0030] Optionally, the calculation of generating the historical context vector includes:

[0031] Numerical weather forecasts from different sources are extracted from the multi-source meteorological data, and the degree of difference between them is calculated to generate a meteorological forecast divergence index.

[0032] Extract the time-of-use electricity price sequence from the electricity market data, calculate its intraday fluctuation range, and generate an electricity price volatility index;

[0033] The historical daily deviations that are similar to the historical scenario vector are analyzed from the power plant operation data, and a historical deviation risk index is calculated and generated in combination with the market assessment rules of the same period.

[0034] The meteorological forecast divergence index, the electricity price volatility index, and the historical deviation risk index are combined into a historical scenario vector.

[0035] Specifically, the meteorological forecast divergence index is calculated. This quantifies the consistency between weather forecasts from different sources, thereby assessing the level of certainty regarding the meteorological conditions on the forecast day. At least three numerical weather forecast sequences for the same historical trading day are automatically retrieved simultaneously from multi-source meteorological data, with key physical quantities being core influencing factors for photovoltaic power generation, such as total horizontal irradiance. Using a 15-minute or 1-hour time granularity, the standard deviation or coefficient of variation between forecast values ​​from different sources is calculated at each time point. Finally, the divergence indices at all time points within the day are time-weighted averaged or directly summed to obtain a comprehensive scalar value, namely the meteorological forecast divergence index.

[0036] Performs calculations of the electricity price volatility index. This measures the intraday volatility of electricity market prices to reveal potential arbitrage opportunities and price risks. It extracts the complete time-of-use (TOU) electricity price sequence for the historical trading day from electricity market data; this sequence is typically 96 or 48 points. The system iterates through this price sequence to identify the highest and lowest electricity prices of the day and calculates the ratio of their difference to the average price for the day, or directly calculates the standard deviation of the price sequence.

[0037] The historical deviation risk index is calculated. Based on the power plant's past performance, the potential deviation assessment losses under similar circumstances are evaluated. The system first searches the historical database for a set of historical days with similar scenarios based on the initially formed meteorological and electricity price characteristics, typically selecting the records of the top 5 to 10 days with the highest similarity.

[0038] The system retrieves reported and actual power generation data for similar days from the power plant's operational data to calculate a time-by-time deviation power sequence. Combining this with market assessment rules from the same period, such as different penalty prices for positive and negative deviations or deviation tolerance thresholds, the system performs retrospective financial settlement on these historical deviation power sequences, simulating the potential deviation assessment costs for each day. Finally, the simulated assessment costs for these similar days are statistically averaged to obtain the historical deviation risk index. This index directly reflects the expected economic losses that the power plant's existing forecasting and reporting models may cause under such circumstances.

[0039] The three independent indices mentioned above are merged into a unified historical context vector. The aim is to form a standardized multi-dimensional feature coordinate system to represent the overall state of that historical trading day in a vector space. The distribution of the context vector space and similarity matching are as follows: Figure 2 As shown in the diagram. This process first normalizes the weather forecast divergence index, electricity price volatility index, and historical deviation risk index, for example, by using a min-max normalization method to map them uniformly to the range of 0 to 1, thus eliminating differences in dimensions and numerical ranges. After normalization, these three standardized indicators are combined as three components of a three-dimensional vector to form the final historical context vector. This process can be expressed by the following formula:

[0040] ,

[0041] in, This represents the final generated historical context vector. The normalized weather forecast divergence index reflects the degree of uncertainty in weather forecasts. The index represents the normalized electricity price volatility intensity, and its value reflects the volatility risk and opportunity of market prices. This represents the normalized historical deviation risk index, whose value reflects the level of deviation assessment loss predicted based on historical experience in this scenario. This vector As a whole, it is stored in the context-policy mapping library and becomes the benchmark for subsequent similarity matching.

[0042] For example, the system performs a recurrence calculation for a specific historical trading day. First, it obtains the numerical weather prediction sequences of three sources for the total horizontal irradiance physical quantity for that day. The predicted values ​​at 12:00 noon are as follows: , and The system calculates the standard deviation at that point in time and performs a time-weighted average of the differences across all points in time during the day to derive the meteorological forecast divergence index. The system then retrieved data from that day containing... The time-of-use electricity price sequence at each moment is used to identify the highest electricity price of the day. Yuan / kWh, the lowest electricity price is The average daily electricity price is [price per kWh]. The price fluctuation index is calculated based on the ratio of the difference between the highest and lowest electricity prices to the average electricity price. The system then retrieves the top results based on similarity. Historical daily records were retrieved, along with data on declared and actual electricity generation, and negative deviation penalties were applied concurrently. The assessment rules based on yuan / kWh were used to retrospectively settle financial accounts, and the statistical average of the simulated assessment costs for similar days was obtained, which is the historical deviation risk index. for Yuan. Finally, the system performs normalization, assuming the meteorological divergence is within the historical range. arrive Inner mapping obtained The intensity of electricity price fluctuations is arrive Inner mapping obtained Historical bias risk arrive Inner mapping obtained Finally, by combining the formulas defined in the instruction manual and substituting the values, the historical context vector for that historical trading day is obtained. .

[0043] Optionally, combining the weather forecast divergence index, the electricity price volatility index, and the historical deviation risk index into a historical scenario vector includes:

[0044] The meteorological forecast divergence index, the electricity price fluctuation intensity index, and the historical deviation risk index are normalized to obtain the normalized index.

[0045] The normalized exponents are weighted and linearly combined to generate a historical context vector.

[0046] Specifically, normalization is performed. The purpose is to eliminate the inherent differences in dimensions and significant variations in numerical ranges among the weather forecast divergence index, electricity price volatility index, and historical deviation risk index. Since these three indices have completely different physical meanings and calculation methods, and the electricity price volatility index is a dimensionless ratio, directly combining them would allow the index with the larger numerical range to dominate subsequent calculations, distorting the true similarity of the scenarios. The system employs a min-max normalization method to process each index.

[0047] A weighted linear combination is performed to generate a historical context vector. The aim is to construct a structured vector from three standardized indices and assign relative importance to different dimensions within this vector space in business decision-making. Here, "weighted linear combination" does not mean simply summing the three indices into a scalar, as that would result in significant loss of dimensional information. Instead, it means using the three normalized indices as components on three orthogonal bases to form a three-dimensional historical context vector, representing a specific point in the three-dimensional context space. The "weighting" is reflected in the application of this vector space. A set of weight coefficients is preset. , , Furthermore, the sum of the three is 1. This set of weights is not used when generating vectors, but rather as a multiplier for the difference in their respective dimensions when calculating the distance between two historical context vectors, thereby adjusting the importance of different dimensions. In this way, the information of each index is fully preserved, and flexible weighting is achieved, which is completely in line with objective laws and avoids the logical problems of directly adding different physical quantities.

[0048] For example, the system performs normalization and weighted linear combination replication operations for a specific historical trading day. The system retrieves the original index data calculated for that day: the weather forecast divergence index is 60.0, the electricity price volatility index is 3.0, and the historical deviation risk index is 4000 yuan. The system performs normalization to eliminate dimensional differences, mapping the weather divergence index (0-100) to a normalized weather index of 0.60 using the min-max normalization method; mapping the electricity price volatility index (0-4.0) to a normalized electricity price index of 0.75; and mapping the historical deviation risk index (0-10000 yuan) to a normalized risk index of 0.40. Subsequently, the system performs a weighted linear combination. This process is not a simple numerical summation, but rather uses the three normalized indices as components on three orthogonal bases, combining them to generate a structured three-dimensional historical context vector. That is, the vector for that day is represented by coordinate points composed of 0.60, 0.75, and 0.40. In subsequent applications, the system pre-sets a set of weighting coefficients to adjust the importance of dimensions, setting the weight for weather forecast divergence at 0.3, electricity price volatility at 0.5, and historical deviation risk at 0.2. When the system needs to calculate the similarity between the historical vector and a predicted day scenario vector with components of 0.50, 0.80, and 0.30, the system multiplies the square of the differences in each dimension by the corresponding weighting coefficient, sums them, and then takes the square root of the sum to obtain the weighted spatial distance, which is approximately 0.079. Finally, the system converts this value into a similarity score using the inverse function of distance; a higher score indicates that the historical day's scenario is closer to the predicted day's scenario.

[0049] S2. Based on the actual operating results of the historical trading days, reverse the derivation of the optimized historical declaration curve, associate the historical scenario vector with the optimized historical declaration curve, and construct a scenario-strategy mapping library;

[0050] Specifically, the system retrieves the entire historical trading day's operational records through a data integration interface. These records include the actual power output of the power plant at each time period, the actual market transaction price, and the detailed calculation rules for the market's power deviation assessment. During the reverse engineering operation, the system uses the actual power generation curve of each historical trading day as a resource boundary constraint and the actual market price sequence as the revenue target weight. A nonlinear optimization algorithm searches the solution space for a simulated declaration scheme that maximizes the final settlement amount for that trading day. This scheme, called the optimized historical declaration curve, represents the most ideal response the power plant can take given all market and physical results. Simultaneously, the system extracts multi-source environmental information for the corresponding date, including solar irradiance measured by a weather station, ambient temperature, and the market's predicted electricity price trend. After data normalization and feature encoding, a historical scenario vector with unique environmental characteristics is formed. Finally, the system binds the historical scenario vector to the corresponding optimized historical declaration curve using an association mapping algorithm and stores it in a distributed storage device, thus constructing a scenario strategy mapping library.

[0051] Optionally, the following steps are included before building the context-policy mapping library:

[0052] A large number of historical context vector samples and corresponding high-quality application curve samples are obtained to form a training set;

[0053] Define an initial neural network model architecture containing an encoder and a decoder to learn the mapping relationship from the context vector to the declaration curve;

[0054] The initial neural network model is trained in a supervised manner using the training set until the model performance converges, resulting in a policy generation model that can generate policy genealogies based on the input context vector.

[0055] Specifically, a training set is constructed. The purpose is to prepare high-quality teaching materials for the neural network to learn from. Long-term historical data is extracted from the historical database. For each historical day, the aforementioned method is first used to calculate its corresponding historical context vector. Simultaneously, based on the actual operating results and market data for that day, the system uses back-optimization to deduce a theoretically optimal high-quality application curve, which typically maximizes the overall return for that day. The system pairs these two elements to form a "context-policy" sample pair. Repeating this process, the system accumulates a large-scale training set, typically requiring thousands of samples to ensure the model's generalization ability.

[0056] Define the initial neural network model architecture. The goal is to design a computational structure capable of effectively processing input context vectors and generating output power time series. The system employs an encoder-decoder architecture. The encoder receives low-dimensional historical context vectors, such as 3 to 15 dimensions, compresses them through several fully connected layers, and maps them into an intermediate state vector, or "thought" vector, that encapsulates the essence of the context. The decoder receives this "thought" vector as initial input and conditional constraints, and through a recurrent neural network unit, such as GRU or LSTM, progressively expands it along the time dimension, sequentially generating the declaration curve for each time segment within the next day.

[0057] Supervised training is performed. Its purpose is to fine-tune and optimize the millions of network parameters within the model by repeatedly learning from the correct answers in the training set, enabling it to "learn by analogy." The system inputs the constructed training set into the initialized neural network model. The training process is performed in batches, with each batch containing, for example, 64 or 128 "context-policy" sample pairs. For each sample, the model generates a predicted declaration curve based on the input context vector. The system then calculates the difference between this predicted curve and the corresponding high-quality declaration curve in the training samples; this difference is quantified by a loss function, typically the mean squared error (MSE).

[0058] ,

[0059] In this formula, It is the loss value for a single sample; This is the total number of time points in the reporting curve, such as 96; Is the model in The predicted power value at any given time; It is the true high-quality application curve in the sample. The system calculates the power value at time step (i.e., the loss value) using optimization algorithms such as Adam. Through backpropagation, it calculates the gradient of the loss with respect to all model parameters and fine-tunes the parameters along the gradient descent direction. This process is repeated hundreds of times on the entire training set until the model's loss on the independent validation set no longer decreases significantly, indicating model performance convergence. After training, this initial neural network model, which has absorbed a wealth of historical wisdom, becomes a policy generation model capable of directly generating high-quality policy families based on the input context vector.

[0060] For example, the system performs reproducible training for building the policy generation model. First, the system accumulates 1000 situational policy sample pairs from two years of historical data to form a training set, which includes historical situational vectors and corresponding high-quality reporting curve samples. The system defines an initial neural network model architecture: the encoder receives a 3D historical situational vector and maps it to an intermediate vector, while the decoder uses recurrent neural network units to generate power values ​​for 96 time segments throughout the day. During the supervised training phase, the system inputs a predicted daily situational vector into the model, assuming the model... Power value predicted at any time The sequence remained at 45MW during the midday period, while the actual high-quality declared power value in the sample was... The system operates at 50MW during that period. The system then calculates the model bias based on the mean squared error loss function defined in the manual. The system then uses an optimization algorithm to backpropagate based on the loss value and fine-tune the weight parameters of the encoder and decoder, repeating the iteration on the entire training set until the model performance converges.

[0061] S3. For the forecast date, obtain the corresponding multi-source meteorological data and power market data, and calculate and generate the forecast date scenario vector;

[0062] Optionally, the step of acquiring the corresponding multi-source meteorological data and electricity market data, and calculating and generating the forecast daily scenario vector includes:

[0063] Feature extraction and standardization are performed on the multi-source meteorological data and electricity market data to obtain standardized feature sequences;

[0064] Principal component analysis and key index fusion are performed on the standardized feature sequence to calculate and generate a predicted daily situation vector.

[0065] Specifically, feature extraction and standardization are performed. The aim is to systematically extract all quantitative information that may affect photovoltaic power generation and market transactions from multi-source meteorological and electricity market data, and transform it into dimensionless, comparable standard numerical sequences. The system first acquires multi-source meteorological data for the forecast day, including but not limited to 15-minute sequences of total horizontal irradiance, temperature, humidity, and wind speed from 3 to 5 different sources, as well as the next day's time-of-use electricity price sequence released by the electricity market. Subsequently, the system performs feature extraction, generating statistical features for each type of time series data, including daily averages, maximum values, minimum values, peak-to-valley differences, standard deviations, and the average of key time periods, such as 11:00 AM to 2:00 PM. For multi-source meteorological data, consistency indicators between predicted values ​​from different sources, such as the mean and standard deviation, are also calculated. This process may generate hundreds of raw features. To eliminate differences in dimensions and numerical ranges between different features, the system standardizes all extracted features.

[0066] Principal component analysis (PCA) and key indicator fusion are performed. The aim is to reduce the dimensionality of the high-dimensional standardized feature sequence, removing information redundancy and noise while retaining the most valuable information for core decision-making, ultimately forming a compact predictive daily scenario vector. The system uses the standardized feature sequence obtained in the previous step as input and applies PCA. PCA is a statistical method that transforms multiple related original features into a few independent composite features, i.e., principal components, through linear transformation. The system selects and retains the top-performing features whose cumulative variance contribution rate reaches a certain threshold, such as 85% to 95%. One principal component. This means that this Principal components can explain most of the variation information in the original data, thus achieving effective information compression. The value of is typically between 5 and 15. Meanwhile, considering that simple statistical dimensionality reduction might overlook certain crucial indicators with clear business implications, the system performs a key indicator fusion operation. The system directly selects several key indicators from the original feature set that have been proven extremely important by expert knowledge or prior models, such as the electricity price volatility index or predicted peak irradiance. These standardized indicators are then concatenated with the principal component vectors obtained after PCA dimensionality reduction to form the final predicted daily scenario vector. Its structure can be represented as:

[0067] ,

[0068] Here, This is the final generated predicted daily scenario vector; This represents the value of the first principal component. This represents the value of the second principal component. Representing the The values ​​of the principal components; The standardized value representing the first selected key indicator. Representing the The standardized values ​​of the selected key indicators.

[0069] For example, the system performs feature extraction and standardization for the forecast day to generate a forecast day scenario vector. First, the system obtains 15-minute total horizontal irradiance sequences from three different sources for the forecast day, as well as the next day's time-of-use electricity price sequence released by the electricity market. The system extracts the intraday peak features from the irradiance sequences, obtaining peak values ​​of 820 W / m², 860 W / m², and 780 W / m² from the three sources, with a mean of 820 W / m² and a standard deviation of 40 W / m². The system then standardizes the extracted features. Assuming the original mean feature value for a key period on the forecast day is 600 W / m², while the corresponding mean in the historical sample set is 500 W / m² and the standard deviation is 50 W / m², the standardization value of this feature is calculated to be 2.0, thus forming a high-dimensional standardized feature sequence. Next, the system uses this sequence as input and applies Principal Component Analysis (PCA) for dimensionality reduction, setting the cumulative variance contribution rate threshold to 90%. The system calculates the preceding... The values ​​of the principal components, assuming The value is 3, resulting in a principal component vector of [0.85, −0.12, 0.45]. Simultaneously, the system performs a key indicator fusion operation, directly selecting the standardized electricity price volatility index. Standardized index for 0.65 and predicted peak irradiance 0.70, these indicators are concatenated with the principal component vector. The system combines them according to the formula defined in the manual, and substituting the above values, yields the final generated forecast daily scenario vector. With a value of [0.85,−0.12,0.45,0.65,0.70], this vector achieves both effective compression of high-dimensional information and ensures the direct representation of key decision-making information.

[0070] S4. Perform similarity matching between the predicted daily context vector and the historical context vector in the context-policy mapping library to retrieve a set of candidate policies;

[0071] Optionally, the candidate strategy set obtained by the retrieval includes:

[0072] Calculate the weighted spatial distance between the predicted daily context vector and each historical context vector in the context-policy mapping library, and convert this distance into a similarity score;

[0073] Based on the similarity scores, the historical context vectors are sorted in descending order, and the top-ranked historical context vectors are selected as the set of similar historical context vectors.

[0074] From the context-policy mapping library, the optimized historical application curves that are uniquely associated with each vector in the set of similar historical context vectors are extracted to form a candidate policy set.

[0075] Specifically, similarity matching calculations are performed. This involves quantifying the similarity between the predicted date and each historical trading day within a comprehensive context. The system uses the generated predicted date context vector as the query benchmark and iterates through each historical context vector stored in the context-strategy mapping library. For each pair of vectors, the system calculates the weighted spatial distance between them. Weighted spatial distance is a metric that allows different importance weights to be assigned to different dimensions of the context vectors. After calculation, the system transforms the resulting distance value into a standardized similarity score using an inverse function. The score typically ranges from 0 to 1, with higher scores indicating stronger similarity. This process can be represented by the following formula:

[0076] ,

[0077] ,

[0078] in, This represents the weighted spatial distance between the predicted daily scenario vector and the i-th historical scenario vector. , , These are preset weights for three dimensions: weather forecast divergence, electricity price volatility, and historical deviation risk. These weights are pre-configured by system operation strategies or expert experience, and their sum is 1. , , These are the three components of the predicted daily situation vector. , , These are the three components of the i-th historical context vector. It is determined by distance The resulting similarity score is inversely proportional to the distance.

[0079] The system filters similar historical scenarios based on similarity scores. This involves precisely identifying a small, highly correlated set of similar days from all historical data. After calculating the similarity of all historical scenario vectors, the system displays the resulting similarity scores. The sequence is sorted in descending order. The selection is based on a preset truncation quantity. Select the top-ranked A number of historical context vectors. This is a key engineering parameter, typically set between 10 and 20, depending on the size of the database and the need for strategy diversity. This selected high-scoring vector subset constitutes a set of similar historical context vectors highly correlated with the prediction date, with the context similarity distribution as follows: Figure 3 As shown.

[0080] Based on the selected similar scenarios, strategy extraction is completed. This involves mapping the identified similar scenarios to specific, actionable application curves. The system takes this set of similar historical scenario vectors as input and performs batch queries using the indexing function of the scenario-strategy mapping library. Since each historical scenario vector in the library is uniquely associated with an optimized historical application curve, the system traverses each similar historical scenario vector in the set and directly extracts its corresponding optimized historical application curve. All these extracted application curves, representing the best practices under similar historical conditions, together form a candidate strategy set, which is then passed to the next processing stage for in-depth analysis and optimization.

[0081] For example, the system performs similarity matching calculations to narrow down candidate strategies. First, the system determines the components of the predicted day context vector, and then obtains the predicted day vector through feature extraction and standardization. for At the same time, a representative historical context vector is selected from the mapping library. for The system pre-configures the weights for each dimension and sets the weights for the degree of divergence in weather forecasts. for Electricity price volatility intensity weighting for Historical bias risk weight for And the sum of weights is The system calculates the weighted spatial distance between the two based on the formula defined in the instruction manual. The system then uses the inverse distance function to convert it into a similarity score. After the system completes the distance calculation for all vectors in the mapping library, it will assign similarity scores. The sequence is sorted in descending order, with a preset number of segments. for The system then selects the highest-ranked... Each historical context vector constitutes a set of similar historical context vectors. Finally, the system uses an indexing function to traverse this set and directly extracts vectors related to these historical context vectors. Each vector is uniquely associated with an optimized historical application curve, thus forming a complete set of candidate strategies.

[0082] S5. Based on each candidate application curve in the candidate strategy set, output a multi-dimensional contradiction diagnosis report containing expected economic benefits and risk indicators;

[0083] Optionally, the output includes a multi-dimensional diagnostic report on the contradictions between expected economic benefits and risk indicators, including:

[0084] For each candidate application curve, perform multiple simulations to obtain a set of deviation power sequences;

[0085] Based on the time-of-use electricity price for the forecast date, the deviation power sequence is divided into a first deviation distribution for high-price periods and a second deviation distribution for low-price periods according to the time of occurrence.

[0086] Based on the first deviation distribution and the preset negative deviation assessment rules, the first risk cost of insufficient power generation during high-yield periods is calculated.

[0087] Based on the second deviation distribution and the market on-grid electricity price, the second resource waste cost of excess power generation during low-yield periods is calculated;

[0088] By integrating the first risk cost with the second resource waste cost, a multi-dimensional contradiction diagnosis report containing expected economic benefits and risk indicators is output.

[0089] Specifically, simulations are performed to generate deviation power sequences. This involves expanding a single, deterministic candidate reporting curve into a set of probabilistic actual power generation results reflecting the uncertainty of weather forecasts. The system initiates a Monte Carlo simulation for a specific candidate reporting curve from the candidate strategy set. Monte Carlo simulation is a computational method that approximates the numerical solution of complex problems through extensive random sampling; here, it is used to simulate the uncertainty of photovoltaic output. Based on multi-source meteorological data for the forecast day, the system constructs a probabilistic photovoltaic power prediction model. This model can output a set of, for example, 200 to 1000 possible actual power generation curves, rather than a single deterministic prediction. The system subtracts each simulated actual power generation curve from the currently evaluated candidate reporting curve at each time segment, thereby generating the same number of deviation power sequences. Each deviation power sequence represents a possible actual deviation scenario, and these sequences together constitute the statistical distribution of deviation power under the reporting strategy.

[0090] Risk stratification calculation is based on economic signals. The goal is to distinguish the different natures and magnitudes of economic impact caused by deviations during different economic value periods. The system first obtains the time-of-use electricity price sequence for the forecast day and sets a high-price threshold. Then, it processes each deviation power sequence generated in the previous step, classifying the deviation values ​​according to their occurrence period. The set of deviation values ​​occurring during high-price periods constitutes the first deviation distribution for high-yield periods; the set of deviation values ​​occurring during low-price periods constitutes the second deviation distribution for low-yield periods.

[0091] The first risk cost is calculated based on the first deviation distribution. This quantifies the direct economic loss caused by a negative deviation—the failure to meet the declared power generation during periods of high electricity prices. The system retrieves preset negative deviation assessment rules, typically formulated by the electricity market operator, which specify the penalty unit price corresponding to different degrees of negative deviation. This unit price is often 1.2 to 2.0 times the market electricity price. For all negative deviation samples in the first deviation distribution, the system calculates the assessment penalty for each sample based on its magnitude and the corresponding electricity price for the time period. Then, it calculates the expected value of the penalty under all simulated scenarios. This expected value is the first risk cost. The calculation formula is as follows:

[0092] ,

[0093] Here, This represents the primary risk cost; [] is the expectation operator, which calculates the average across all simulation scenarios; It is a period of high electricity prices Within a given timeframe; Is The absolute value of the negative deviation power occurring at any given moment; yes The market electricity price at any given time; It is the penalty coefficient defined in the negative deviation assessment rules.

[0094] The second resource waste cost is calculated based on the second deviation distribution. This quantifies the opportunity cost of resource waste caused by exceeding the declared value during periods of low electricity prices, i.e., a positive deviation. In many electricity markets, excess power generation can only be settled at a lower, fixed market feed-in tariff, and may even be subject to uncompensated curtailment. The system multiplies all positive deviation samples in the second deviation distribution by the market feed-in tariff to calculate the low revenue from this excess power generation. This revenue itself is considered a resource waste cost because it occupies power generation capacity that could generate higher value during periods of high prices. The calculation formula is as follows:

[0095] ,

[0096] in, This is the second type of resource waste cost; During periods of low electricity prices Inside The positive deviation power value occurring at any given moment; It is the market-based electricity price, which is usually significantly lower than the market average electricity price.

[0097] The system integrates the calculated primary risk cost and secondary resource waste cost, along with the expected theoretical benefits calculated based on candidate application curves and predicted electricity prices, into a structured, multi-dimensional diagnostic report.

[0098] For example, the system generates a multi-dimensional contradiction diagnosis report for a specific candidate application curve. First, the system initiates a Monte Carlo simulation, constructs a probabilistic photovoltaic power prediction model based on multi-source meteorological data for the forecast date, and performs 500 random samplings. The system then subtracts the 500 simulated actual power generation curves from the candidate application curve at 96 time sections, resulting in a set of deviation power sequences containing 500 records. The system subsequently obtains the time-of-use electricity price for the forecast date, defining the 25% period with the highest electricity price as the high-price period. The rest are periods with low electricity prices. The system categorizes the deviation power sequence by time period, obtaining a first deviation distribution for high-yield periods and a second deviation distribution for low-yield periods. For the first deviation distribution, it is assumed that during periods of high electricity prices… At a certain moment The absolute value of the negative deviation power generated by 500 simulations The average value is 2MW. The system retrieves the preset negative deviation assessment rules, sets the penalty coefficient to 1.5, and sets the market electricity price. The cost during this period is 0.8 yuan / kWh. The expected value of the penalties for all negative deviation samples during this period is calculated, and then substituted into the formula to determine the first risk cost. for Yuan. Regarding the second bias distribution, it is assumed that during periods of low electricity prices... At a certain moment Positive deviation power value The simulated average is 3MW, and the system obtains the market grid-connected electricity price. The cost is 0.3 yuan / kWh. The expected resource waste is calculated based on the positive deviation power value for this period. Substituting this value into the formula, and if this period includes the remaining 72 cross-sections, the second resource waste cost is calculated. for Finally, the system integrates the expected theoretical benefits, primary risk costs, and secondary resource waste costs of the candidate application curve, and outputs a structured, multi-dimensional contradiction diagnosis report that includes expected economic benefits and risk indicators.

[0099] S6. Based on the strategy contradictions revealed in the multi-dimensional contradiction diagnosis report, replace and reorganize the specific time segment of each candidate application curve to generate a derivative strategy set.

[0100] Optionally, the set of generation derivative strategies includes:

[0101] Determine whether the first risk cost in the multi-dimensional contradiction diagnosis report exceeds the preset first risk threshold. If so, select a curve segment with a more conservative prediction value during the corresponding high electricity price period from the candidate strategy set, replace the original curve segment, and generate the first type of derivative strategy.

[0102] Determine whether the second resource waste cost in the multi-dimensional contradiction diagnosis report exceeds the preset second opportunity cost threshold. If so, select the curve segment with a more optimistic prediction value in the corresponding low electricity price period from the candidate strategy set, replace the original curve segment, and generate a second type of derivative strategy.

[0103] All the first type of derivative strategies and the second type of derivative strategies are combined to form a derivative strategy set.

[0104] Specifically, the system generates the first type of derivative strategy. This aims to proactively reduce the risk of hefty penalties due to insufficient power generation during periods of high electricity prices. The system first sets a dynamically adjustable first risk threshold, typically between 1% and 5% of the power plant's average daily revenue, representing the maximum negative deviation risk cost the decision-maker is willing to bear. The system then examines the multi-dimensional contradiction diagnosis report corresponding to each candidate declaration curve. When the first risk cost of a curve exceeds the preset first risk threshold, a derivative operation is triggered. The system identifies the main period causing this high risk, namely, the high electricity price period. Subsequently, it searches the entire candidate strategy set for curve segments with relatively more conservative declared power values ​​(i.e., lower values) during these high electricity price periods. Conservative curve segments refer to those whose declared values ​​are closer to the low quantiles of the photovoltaic power generation probability distribution, such as the P10 or P25 quantile, thus having a higher probability of achieving or exceeding the declared value. The segments in the original candidate declaration curves corresponding to the high electricity price period are replaced with these more conservative curve segments found from other candidate strategies, thereby generating a new declaration curve, i.e., the first type of derivative strategy.

[0105] A second type of derivative strategy is generated to reduce overcapacity and resource waste caused by overly conservative reporting during low-electricity-price periods. The system sets a second opportunity cost threshold, which represents the upper limit of tolerable resource waste, typically estimated based on the difference between the feed-in tariff and the average market tariff. When the system detects that the second resource waste cost of a candidate reporting curve exceeds this threshold in its multi-dimensional contradiction diagnosis report, the generation of a second type of derivative strategy is triggered. Unlike the generation logic of the first type of strategy, the system identifies low-electricity-price periods that cause higher resource waste costs. Then, in the candidate strategy set, the system searches for curve segments with more optimistic reported power values ​​during these low-electricity-price periods, i.e., higher values. Optimistic curve segments mean that their reported values ​​are closer to the high quantiles of the photovoltaic power generation probability distribution, such as the P75 or P90 quantile. The system replaces these low-electricity-price period segments in the original curves with these found more optimistic curve segments, forming a second type of derivative strategy.

[0106] All generated first-class and second-class derived strategies are aggregated. Both types of strategies are local optimizations based on existing candidate strategies. They inherit the rationality of the original strategies for most periods while specifically addressing their most prominent economic contradictions. This set of all newly generated strategies constitutes the derived strategy set.

[0107] For example, the system performs reproducibility optimization for a specific candidate application curve. First, the system retrieves the multi-dimensional contradiction diagnosis report for that curve, identifying its first risk cost during periods of high electricity prices as... Yuan, while the preset first risk threshold is Yuan. Because the first risk cost exceeds the threshold, the system triggers the first type of derivative strategy generation logic, retrieving a segment from the candidate strategy set with a more conservative declared power value during this high electricity price period. The original segment's power is... The power of the retrieved conservative fragment is The system executes fragment replacement to generate the first type of derivative strategy to reduce the risk of negative bias penalty. Subsequently, the system monitors that the second resource waste cost of this curve during low electricity price periods is... The amount exceeded the preset second opportunity cost threshold. Yuan. The system then triggers the second type of derivative strategy generation logic, searching the set for segments with more optimistic reported power values ​​within that time period, and then converting the original... Replace the conservative segment with The system generates a second type of derivative strategy based on optimistic segments to capture more electricity revenue from grid connection. Finally, the system combines the first type of derivative strategy (containing risk hedging segments) with the second type of derivative strategy (containing return enhancement segments) to form a complete set of derivative strategies. This set, through refined restructuring over local time periods, ensures that the application plan can dynamically adapt to the complex electricity price fluctuation environment.

[0108] S7. Score each candidate application curve in the candidate strategy set and the derived strategy set, and select the candidate application curve with the higher score as the target application strategy.

[0109] Specifically, the system iterates through each candidate application curve in the candidate strategy set and the derived strategy set, determining its final score by quantitatively calculating the comprehensive weighted value of various evaluation indicators. First, the system extracts the similarity score corresponding to each curve, representing the degree of fit between the historical context associated with the strategy and the current forecast day's weather and market environment. Next, the system calculates the expected settlement revenue for each curve, which is the cumulative sum of the product of the applied electricity volume and the predicted time-of-use electricity price for each time period, reflecting the economic potential of the strategy under ideal forecast conditions. To measure the robustness of the strategy in actual implementation, the system introduces a first risk cost and a second resource waste cost. The first risk cost refers to the potential loss of revenue or performance penalties due to insufficient actual power generation during high-electricity-price periods. The second resource waste cost refers to the cost of inefficient resource utilization due to excessive application during low-electricity-price periods. During performance evaluation, the system subtracts the sum of the above two costs from the expected settlement revenue and adds the similarity score after conversion by a revenue conversion factor. In this logical operation, all indicators are unified into monetary value units through coefficient processing, thus ensuring that the addition and subtraction operations conform to objective physical logic. Finally, the system sorts the scores of all curves, eliminates the schemes with lower scores, and selects the candidate application curve with the highest score as the final target application strategy.

[0110] S8. After the trading day ends, obtain actual market feedback data containing actual settlement results, and update the scenario-strategy mapping library and the strategy generation model used to generate the target declaration strategy based on the actual market feedback data.

[0111] Optionally, updating the context-policy mapping library and the policy generation model used to generate the target declaration policy includes:

[0112] Based on the actual market feedback data, the actual comprehensive return of the target application strategy on the forecast date is calculated;

[0113] Based on the complete market information and actual power generation data for the predicted date, the optimized post-event strategy for that day is calculated in reverse optimization.

[0114] Associate the predicted daily scenario vector with the optimized post-event strategy to form a new mapping pair and store it in the scenario-policy mapping library;

[0115] The newly added mapping pairs are used as training samples to incrementally train the policy generation model, optimizing its ability to generate target declaration strategies based on context vectors.

[0116] Specifically, the actual comprehensive benefit calculation is performed. This involves accurately quantifying the final economic performance of the implemented target declaration strategy under real market conditions. After the trading day ends, the system automatically obtains actual market feedback data containing the actual settlement results from the market operator. This data package is usually released on the day after the trading day, i.e., T+1 day, and includes the final settlement electricity price for each period, as well as the actual assessment fee or reward amount for power generation deviation. The system aligns this data with the power plant's own actual power generation data and calculates the actual comprehensive benefit of the target declaration strategy on the forecast day according to the following formula:

[0117] ,

[0118] In this formula, For actual comprehensive benefits; Iterate through all settlement periods within a trading day; yes Actual power generation during the period; yes The actual settlement price for the time period; yes The actual deviation power during the time period, that is, the difference between the actual power generation and the target declared value; It is the deviation penalty cost function defined by market rules; It is the deviation reward function defined by market rules.

[0119] Reverse optimization calculates and optimizes post-event strategies. It aims to find the perfect reporting strategy that theoretically maximizes revenue on the predicted date, based on completely certain information, providing an ideal learning benchmark for machine learning models. At this point, complete market information for the predicted date, including actual time-of-use electricity prices, market assessment rules, and the actual power generation capacity of the power plant (i.e., actual power generation data), has become deterministic input. The system constructs an optimization problem whose objective function is to maximize the actual comprehensive revenue in the above formula. The decision variables are the power points along the entire declaration curve. This is a nonlinear programming problem under constraints, such as the declared power not exceeding the power plant's nameplate capacity. Solving this optimization problem yields the optimized post-event strategy for that day.

[0120] The scenario-strategy mapping library is updated. This involves solidifying the complete experience of the current trading day into a new knowledge record to expand the system's historical experience base. The system pairs the scenario vector generated before the trading day with the optimized post-trade strategy calculated via reverse optimization to form a new mapping pair. This mapping pair represents the scenario and strategy that is proven to be the optimal solution in hindsight. Subsequently, the system stores this new mapping pair as a new record in the scenario-strategy mapping library database.

[0121] Incremental training is performed on the policy generation model. The goal is to internalize newly acquired "perfect experience" into the model's capabilities, thereby improving its generalization ability. The newly generated mapping pair from the previous step is used as a new training sample. The policy generation model, typically a deep neural network, undergoes an incremental training process. Incremental training is an efficient model update method that fine-tunes an already trained model using a small amount of new data, without having to train from scratch. In this process, the model takes the predicted daily situation vector as input, attempts to generate a reporting curve, and aims to optimize the post-event policy as output. The network weights are fine-tuned using backpropagation to reduce the difference between the generated curve and the optimized post-event policy. This training process typically uses a small learning rate to ensure that the integration of new knowledge does not compromise the model's existing generalization ability.

[0122] For example, the system performs closed-loop self-optimization and replication calculations during the feedback learning phase after the end of the trading day. First, the system obtains feedback data from the market operator, including actual settlement results, such as the actual power generation of the target reporting strategy during each settlement period on a given forecast day. The average is 40MW, corresponding to the actual settlement price of electricity. The average price is 0.5 yuan / kWh, and the actual deviation power is... The resulting deviation penalty cost The amount is 2000 yuan. Since there is no bonus, the deviation bonus income is... The cost is 0 yuan. The system calculates based on the formula defined in the instruction manual. If there are a total of 96 cross-sections throughout the day (24 hours), the actual comprehensive revenue is... The cost is 40MW × 1000 × 0.5 × 24 - 2000 = 478,000 yuan. Subsequently, the system uses the complete market electricity price, assessment rules, and the actual generating capacity of 40MW as deterministic inputs on the forecast date to construct a system that maximizes... For a nonlinear programming problem with a target, backpropagation derives an optimal ex-post strategy curve that theoretically avoids all penalties and maximizes the capture of high-price periods. The system then pairs the predicted daily scenario vector generated before the transaction with this optimized ex-post strategy curve, forming a new mapping pair that reflects the relationship between the scenario and the optimal decision, and stores it in a scenario-policy mapping library. Finally, the system uses this new mapping pair as training samples, incrementally training the initial neural network model with a small learning rate. Through backpropagation, the weights of the encoder and decoder are fine-tuned, enabling the model to generate predicted values ​​that more closely approximate the optimized ex-post strategy when given the same scenario vector, thus achieving dynamic evolution of the strategy generation model.

[0123] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides a next-day photovoltaic power generation optimization forecasting system for electricity market transactions, the system comprising:

[0124] The data acquisition and processing module is used to acquire multi-source meteorological data, power market data and power plant operation data corresponding to historical trading days, and to calculate and generate historical context vectors based on the acquired data;

[0125] The scenario strategy library construction module is used to reverse-engineer the optimized historical declaration curve based on the actual operating results of the historical trading days, associate the historical scenario vector with the optimized historical declaration curve, and construct a scenario-strategy mapping library.

[0126] The scenario vector calculation module is used to obtain the corresponding multi-source meteorological data and power market data for the forecast date, and calculate and generate the scenario vector for the forecast date.

[0127] The similarity scenario matching and retrieval module is used to perform similarity matching between the predicted daily scenario vector and the historical scenario vector in the scenario-policy mapping library, and retrieve a set of candidate policies.

[0128] The multi-dimensional contradiction diagnosis module is used to output a multi-dimensional contradiction diagnosis report containing expected economic benefits and risk indicators based on each candidate application curve in the candidate strategy set.

[0129] The strategy derivation and optimization module is used to replace and reorganize specific time segments of each candidate application curve based on the strategy contradictions revealed in the multi-dimensional contradiction diagnosis report, and generate a set of derived strategies.

[0130] The strategy evaluation and decision-making module is used to score each candidate application curve in the candidate strategy set and the derived strategy set, and select the candidate application curve with the higher score as the target application strategy.

[0131] The feedback learning and model update module is used to obtain actual market feedback data containing actual settlement results after the end of the trading day, and update the scenario-strategy mapping library and the strategy generation model used to generate the target declaration strategy based on the actual market feedback data.

[0132] 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.

[0133] 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 optimizing and predicting next-day photovoltaic power generation for electricity market transactions, characterized in that, The method includes: Acquire multi-source meteorological data, electricity market data, and power plant operation data corresponding to historical trading days, and calculate and generate historical scenario vectors based on the acquired data; Based on the actual operating results of the historical trading days, the optimized historical declaration curve is deduced in reverse, and the historical scenario vector is associated with the optimized historical declaration curve to construct a scenario-strategy mapping library; For the forecast date, acquire the corresponding multi-source meteorological data and electricity market data, and calculate and generate the forecast date scenario vector; The predicted daily context vector is matched with the historical context vector in the context-policy mapping library to retrieve a set of candidate policies; Based on each candidate application curve in the candidate strategy set, a multi-dimensional contradiction diagnosis report containing expected economic benefits and risk indicators is output. Based on the strategic contradictions revealed in the multi-dimensional contradiction diagnosis report, specific time segments of each candidate application curve are replaced and reorganized to generate a derivative strategy set; Each candidate application curve in the candidate strategy set and the derived strategy set is scored, and the candidate application curve with the higher score is selected as the target application strategy. After the trading day ends, obtain actual market feedback data containing the actual settlement results, and update the scenario-strategy mapping library and the strategy generation model used to generate the target declaration strategy based on the actual market feedback data.

2. The method for optimizing and predicting next-day photovoltaic power generation for electricity market transactions according to claim 1, characterized in that, The calculation to generate the historical context vector includes: Numerical weather forecasts from different sources are extracted from the multi-source meteorological data, and the degree of difference between them is calculated to generate a meteorological forecast divergence index. Extract the time-of-use electricity price sequence from the electricity market data, calculate its intraday fluctuation range, and generate an electricity price volatility index; The historical daily deviations that are similar to the historical scenario vector are analyzed from the power plant operation data, and a historical deviation risk index is calculated and generated in combination with the market assessment rules of the same period. The meteorological forecast divergence index, the electricity price volatility index, and the historical deviation risk index are combined into a historical scenario vector.

3. The method for optimizing and predicting next-day photovoltaic power generation for electricity market transactions according to claim 1, characterized in that, The candidate strategy set obtained by the retrieval includes: Calculate the weighted spatial distance between the predicted daily context vector and each historical context vector in the context-policy mapping library, and convert this distance into a similarity score; Based on the similarity scores, the historical context vectors are sorted in descending order, and the top-ranked historical context vectors are selected as the set of similar historical context vectors. From the context-policy mapping library, the optimized historical application curves that are uniquely associated with each vector in the set of similar historical context vectors are extracted to form a candidate policy set.

4. The method for optimizing and predicting next-day photovoltaic power generation for electricity market transactions according to claim 1, characterized in that, The output includes a multi-dimensional diagnostic report on the contradictions between expected economic benefits and risk indicators, including: For each candidate application curve, perform multiple simulations to obtain a set of deviation power sequences; Based on the time-of-use electricity price for the forecast date, the deviation power sequence is divided into a first deviation distribution for high-price periods and a second deviation distribution for low-price periods according to the time of occurrence. Based on the first deviation distribution and the preset negative deviation assessment rules, the first risk cost of insufficient power generation during high-yield periods is calculated. Based on the second deviation distribution and the market on-grid electricity price, the second resource waste cost of excess power generation during low-yield periods is calculated; By integrating the first risk cost with the second resource waste cost, a multi-dimensional contradiction diagnosis report containing expected economic benefits and risk indicators is output.

5. The method for optimizing and predicting next-day photovoltaic power generation for electricity market transactions according to claim 1, characterized in that, The set of generation derivative strategies includes: Determine whether the first risk cost in the multi-dimensional contradiction diagnosis report exceeds the preset first risk threshold. If so, select a curve segment with a more conservative prediction value during the corresponding high electricity price period from the candidate strategy set, replace the original curve segment, and generate the first type of derivative strategy. Determine whether the second resource waste cost in the multi-dimensional contradiction diagnosis report exceeds the preset second opportunity cost threshold. If so, select the curve segment with a more optimistic prediction value in the corresponding low electricity price period from the candidate strategy set, replace the original curve segment, and generate a second type of derivative strategy. All the first type of derivative strategies and the second type of derivative strategies are combined to form a derivative strategy set.

6. The method for optimizing and predicting next-day photovoltaic power generation for electricity market transactions according to claim 1, characterized in that, The updating of the context-policy mapping library and the policy generation model used to generate the target declaration policy includes: Based on the actual market feedback data, the actual comprehensive return of the target application strategy on the forecast date is calculated; Based on the complete market information and actual power generation data for the predicted date, the optimized post-event strategy for that day is calculated in reverse optimization. Associate the predicted daily scenario vector with the optimized post-event strategy to form a new mapping pair and store it in the scenario-policy mapping library; The newly added mapping pairs are used as training samples to incrementally train the policy generation model, optimizing its ability to generate target declaration strategies based on context vectors.

7. The method for optimizing and predicting next-day photovoltaic power generation for electricity market transactions according to claim 1, characterized in that, The process of acquiring the corresponding multi-source meteorological data and electricity market data, and calculating and generating the forecast daily scenario vector includes: Feature extraction and standardization are performed on the multi-source meteorological data and electricity market data to obtain standardized feature sequences; Principal component analysis and key index fusion are performed on the standardized feature sequence to calculate and generate a predicted daily situation vector.

8. The method for optimizing and predicting next-day photovoltaic power generation for electricity market transactions according to claim 2, characterized in that, The step of combining the weather forecast divergence index, the electricity price volatility index, and the historical deviation risk index into a historical scenario vector includes: The meteorological forecast divergence index, the electricity price fluctuation intensity index, and the historical deviation risk index are normalized to obtain the normalized index. The normalized exponents are weighted and linearly combined to generate a historical context vector.

9. The method for optimizing and predicting next-day photovoltaic power generation for electricity market transactions according to claim 1, characterized in that, Before building the context-policy mapping library, the following also includes: A large number of historical context vector samples and corresponding high-quality application curve samples are obtained to form a training set; Define an initial neural network model architecture containing an encoder and a decoder to learn the mapping relationship from the context vector to the declaration curve; The initial neural network model is trained in a supervised manner using the training set until the model performance converges, resulting in a policy generation model that can generate policy genealogies based on the input context vector.

10. A next-day photovoltaic power generation optimization forecasting system for electricity market transactions, applied to the next-day photovoltaic power generation optimization forecasting method for electricity market transactions as described in any one of claims 1-9, characterized in that, The system includes: The data acquisition and processing module is used to acquire multi-source meteorological data, power market data and power plant operation data corresponding to historical trading days, and to calculate and generate historical context vectors based on the acquired data; The scenario strategy library construction module is used to reverse-engineer the optimized historical declaration curve based on the actual operating results of the historical trading days, associate the historical scenario vector with the optimized historical declaration curve, and construct a scenario-strategy mapping library. The scenario vector calculation module is used to obtain the corresponding multi-source meteorological data and power market data for the forecast date, and calculate and generate the scenario vector for the forecast date. The similarity scenario matching and retrieval module is used to perform similarity matching between the predicted daily scenario vector and the historical scenario vector in the scenario-policy mapping library, and retrieve a set of candidate policies. The multi-dimensional contradiction diagnosis module is used to output a multi-dimensional contradiction diagnosis report containing expected economic benefits and risk indicators based on each candidate application curve in the candidate strategy set. The strategy derivation and optimization module is used to replace and reorganize specific time segments of each candidate application curve based on the strategy contradictions revealed in the multi-dimensional contradiction diagnosis report, and generate a set of derived strategies. The strategy evaluation and decision-making module is used to score each candidate application curve in the candidate strategy set and the derived strategy set, and select the candidate application curve with the higher score as the target application strategy. The feedback learning and model update module is used to obtain actual market feedback data containing actual settlement results after the end of the trading day, and update the scenario-strategy mapping library and the strategy generation model used to generate the target declaration strategy based on the actual market feedback data.

Images