Photovoltaic power station point-probability joint prediction adaptive switching method based on reinforcement learning

Abstract

The invention discloses a photovoltaic power station point-probability joint prediction adaptive switching method based on reinforcement learning, and aims to solve the contradiction that point prediction under high-proportion photovoltaic access is not accurate on a fluctuation day and the probability prediction cost is high. The method comprises the following steps: extracting a normalized output track and a climbing intensity feature based on photovoltaic output data of a previous day, and constructing a decision state without meteorological dependence in combination with calendar prior; a deep reinforcement learning network is adopted to learn a mapping relation of a previous day form-a next day fluctuation risk-prediction mode selection, and an automatic switching action of point prediction and probability prediction is output; and designing a reward function fusing the fluctuation index and the prediction cost, driving the strategy to preferentially select point prediction on a typical day, and automatically switching to probability prediction on a fluctuation day. According to the method, collaborative optimization among prediction precision, robustness and calculation cost is realized, and the adaptability and economy of a photovoltaic prediction system under complex weather and multi-site conditions are improved.

Description

Technical Field

[0001] This invention belongs to the field of new energy power prediction and power system operation control technology, and specifically relates to a photovoltaic power station-probabilistic joint prediction adaptive switching method based on reinforcement learning. Background Technology

[0002] As the penetration rate of photovoltaic power generation in distribution networks, microgrids, and regional power grids continues to increase, output uncertainty has evolved from a "negligible disturbance" to a key constraint affecting dispatch plans, reserve configuration, energy storage deployment, and market clearing. The fluctuations in photovoltaic output mainly stem from rapid changes in irradiance caused by cloud cover, aerosol variations, and local convection, exhibiting significant time-varying and non-stationarity. This results in the same site displaying two distinct output patterns on different dates: "typical days (regular and smooth changes)" and "fluctuating days (frequent ramp-ups and significant abrupt changes)."

[0003] Existing photovoltaic power prediction engineering systems typically employ point prediction models (such as single-valued regression based on statistical learning or deep learning) and probabilistic prediction models (such as quantile regression, scene generation, or Bayesian inference). Point prediction models offer advantages such as lightweight structure, fast inference, and ease of integration; however, errors are often significantly amplified on fluctuating days, and these errors exhibit asymmetric risk (both underestimation and overestimation can have adverse consequences for scheduling). Probabilistic prediction models, on the other hand, provide a characterization of uncertainty and are more robust on fluctuating days; however, they generally require more complex modeling, higher inference overhead, and greater operational costs (training, calibration, storage, and post-processing).

[0004] However, existing methods generally have the following shortcomings:

[0005] (1) The prediction model lacks an adaptive selection mechanism. Most systems lack interpretable and learnable switching logic between point prediction and probability prediction. They usually adopt fixed strategies or manual rules, which are difficult to adapt to seasonal changes and changes in the proportion of fluctuating days.

[0006] (2) The decision-making stage is highly dependent on future information. Many switching rules rely on future weather forecasts, cloud images or irradiance predictions, and are prone to failure when meteorological data is missing or of unstable quality.

[0007] (3) Deep coupling between point / probability predictors leads to engineering complexity. Embedding the switching logic into the prediction model often results in a long training chain, difficulty in parameter tuning, and high cost of online maintenance, making it difficult to achieve "lightweight deployment + robust operation".

[0008] Therefore, there is an urgent need for a technical solution that does not rely on future meteorological information and can automatically identify "next-day fluctuation risk" and output the selection of prediction models based solely on historical operational data, so as to reduce the overall system cost and improve deployability while ensuring robustness. Summary of the Invention

[0009] This invention aims to provide a non-predictive reinforcement learning switching strategy method for photovoltaic power prediction services. It is used to automatically decide whether to use point prediction or probability prediction for the next day based on the power generation pattern of the previous day on a daily scale. This solves the problems of insufficient robustness of point prediction on fluctuating days in high-penetration photovoltaic scenarios, high cost and difficulty in full-scale application of probability prediction, strong dependence of switching strategies on future meteorological information, and high engineering coupling.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0011] A reinforcement learning-based photovoltaic power station-probabilistic joint prediction adaptive switching method includes the following steps:

[0012] S1. Historical Photovoltaic Power Output Data Acquisition and Intraday Sample Construction: When the photovoltaic power station is in normal operation or in a state awaiting prediction, acquire the historical power output time series data of the photovoltaic power station. The historical power output time series data includes a timestamp and the photovoltaic active power sequence of multiple stations. Reconstruct the historical power output time series data according to a preset intraday time window to form a daily power station output matrix. ;

[0013] S2. Extraction of morphological features and construction of state variables from the previous day: Extracting the previous day's force matrix. Normalization and differencing are performed to extract characteristic tensors representing the force pattern and wave properties. The feature tensor includes normalized output trajectory features and climbing intensity features; simultaneously, the calendar prior feature vector of the target day d is extracted. The calendar prior feature vector Including the sine and cosine characteristics of the position within the year and the characteristics of the week; and Together they constitute a state of no predictive decision-making. ;

[0014] S3. Construction and Output of the Switching Policy Network Based on Deep Reinforcement Learning: Construct a Deep Q-Network (DQN) as the switching policy network. The DQN includes components for processing... The convolutional neural network branch and the branch used for processing The multilayer perceptron branch outputs an action value vector at the fusion layer. based on Greedy strategy or greedy strategy selection action And output the binary switching result, where Indicates the target date Using a point prediction model, Indicates the target date A probabilistic prediction model is used;

[0015] S4. Reward Function Construction and Offline Training: Based on the Actual Output Curve of the Target Day d Calculate volatility index And based on the volatility distribution of the training set, a site threshold is constructed for each site. ;according to With site threshold Calculate the substitution loss function between point prediction cost and probability prediction cost, and take into account the cost of probability prediction to form the reward. Used to update DQN parameters; the DQN is trained offline through experience replay and target network mechanism to obtain the training switching strategy;

[0016] S5. Automatic strategy output and prediction process separation for validation sets: Generate status for the validation set daily in chronological order. The trained switching strategy is invoked to output the switching results for each target day and write them into the strategy file. Based on the strategy file, the prediction task for the target day is automatically diverted to the point prediction model or the probability prediction model for execution, generating the corresponding point prediction results or probability prediction results.

[0017] Furthermore, the sub-steps of S1 are as follows:

[0018] S1.1: Collect hourly photovoltaic power data by site and align and deduplicate the timestamps;

[0019] S1.2: Select sample points from 06:00 to 18:00 each day and resample them to keep the number of sample points fixed at 13 per day;

[0020] S1.3: Perform interpolation and imputation processing on missing points within the day. The interpolation is preferably linear interpolation or spline interpolation based on time index.

[0021] S1.4: Forming a daily stacked sample tensor , where N is the number of sample days and 13 is the number of sampling points within a day.

[0022] Furthermore, the aforementioned photovoltaic power station-probabilistic joint prediction adaptive switching method is characterized by the following sub-step in S2:

[0023] S2.1: For the previous day's power matrix Normalize the trajectory features by the maximum value or quantile at each station to obtain the normalized trajectory features. ;

[0024] S2.2: For the normalized trajectory features Calculate the difference between adjacent time points and take the absolute value to obtain the climbing intensity characteristics. ;

[0025] S2.3: Will and Formation of morphological feature tensors by stacking channels ;

[0026] S2.4: Calculate the calendar prior features of the target date The calendar prior features Including positions within the year , And the one-hot coding features of the week;

[0027] S2.5: Will and Together they constitute a state of no predictive decision-making. .

[0028] Furthermore, the processing steps for the action value vector output by the Deep Q-Network (DQN) described in S3 include:

[0029] S3.1: Convolutional branches use at least three convolutional layers. Feature extraction is performed, and a batch normalization layer and a non-linear activation function are set after each convolutional layer to obtain the image embedding vector.

[0030] S3.2: Prior branches use at least two fully connected layers. Perform feature mapping to obtain the prior embedding vector;

[0031] S3.3: Concatenate the image embedding vector with the prior embedding vector and input it into the fusion fully connected layer to output the action value vector. , where K is the dimension of the action space.

[0032] Furthermore, the action space of S3.3 is one of the following:

[0033] (1) A unified action space for the entire site. The corresponding output is {0,1};

[0034] (2) Multi-site differentiated action space, The action 'a' encodes the site switching result in the form of a bitmask, and the j-th bit of 'a' indicates whether the j-th site uses point prediction or probability prediction on the target day.

[0035] Furthermore, the volatility index in S4 The construction includes one or a combination of the following:

[0036] (1) The ratio of the average absolute climbing intensity to its relative mean based on the power difference between adjacent times within the day;

[0037] (2) Based on the ratio of the standard deviation of the intraday power series to its relative mean;

[0038] (3) Morphological fluctuation indicators based on intraday peak-to-valley difference, climbing extremes, or the number of multiple peaks:

[0039]

[0040] in, Indicates the first The first day A daily segment or statistics window, Indicates the switching strategy. For policy network parameters, Represents the expectation operator. This represents the loss function.

[0041] Furthermore, the site threshold in S4 The construction method is as follows: quantile thresholds are calculated for each station of the training set volatility samples. ,in The preset quantile parameter is used to distinguish between typical days and fluctuating days. Used to delineate sites Typical fluctuation boundaries; during global handover, the thresholds of each site are aggregated into a site threshold through a weighted average. .

[0042] Furthermore, in S4, the reward r(d) is constructed using an alternative loss function, which includes:

[0043] (1) Point prediction substitution loss ;

[0044] (2) Probability prediction substitution loss ;

[0045] in Use cost terms for probabilistic prediction. For preset coefficients, These represent the basic cost of probabilistic prediction and the sensitivity coefficient of probabilistic prediction to out-of-threshold fluctuations, respectively. These represent the base cost of point prediction and the slope of the point prediction over-threshold penalty, respectively.

[0046] Reward definition .

[0047] Furthermore, S4's offline training employs a deep Q-learning algorithm, including:

[0048] (1) Adopt Greedy strategy selects actions and gradually decays them. To achieve a balance between exploration and utilization;

[0049] (2) Use an experience replay mechanism to randomly select a small batch of samples from historical interaction samples for parameter updates;

[0050] (3) Use a target network mechanism or a dual DQN mechanism to calculate the target Q value to improve training stability;

[0051] (4) Minimize the Q-value estimation error by using Huber loss or mean squared error loss, and update the network parameters by using Adam optimizer.

[0052] Furthermore, the automatic policy output for the validation set in S5 includes: using the validation set... The day before yesterday, the power matrix Build status The switching results are output by the trained DQN and written to the policy file to drive the automatic splitting of subsequent point prediction and probability prediction processes.

[0053] This invention significantly improves the adaptability, economy, and engineering deployability of photovoltaic power prediction systems under complex weather and multi-site operation conditions by integrating "morphological perception, prediction-free decision-making, and cost constraints," providing replicable key support for new energy power prediction and scheduling decisions. Attached Figure Description

[0054] Figure 1 This is a flowchart of a photovoltaic power station-probability joint prediction adaptive switching method based on reinforcement learning, according to an embodiment of the present invention. Detailed Implementation

[0055] 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.

[0056] like Figure 1 As shown, the first aspect of this invention provides a photovoltaic power station-probabilistic joint prediction adaptive switching method based on reinforcement learning, the steps of which are as follows:

[0057] S1: Historical PV Output Data Acquisition and Intraday Sample Construction: Aiming to construct standardized samples suitable for daily-scale decision-making, this involves collecting or retrieving data containing timestamps. Historical photovoltaic power data from multiple sites is collected; the data is reconstructed according to a preset intraday time window, which is from 06:00 to 18:00 daily, with a sampling frequency of once per hour, so that each day forms an intraday power sequence of fixed length 13, and stacked to form a daily sample tensor. ,in For the number of days, The number of sites is used; missing points within the day are interpolated and imputed to ensure sample integrity and trainability.

[0058] S2: Previous Day's Morphological Features Extraction and State Quantity Construction: Aiming to characterize the observable features of "next day's volatility risk" without introducing future information, this involves the previous day's power matrix. Normalization is performed to construct normalized force trajectory characteristics. ;right By performing the difference between adjacent time points and taking the absolute value, the climbing intensity characteristic can be constructed. ;Will and Formation of morphological feature tensors by stacking channels Simultaneously, construct the calendar prior feature vector of the target day d. The calendar prior feature vector It should include at least the sine and cosine characteristics of the position within the year and the one-hot encoding characteristics of the week; Together they constitute a state of no predictive decision-making. ;

[0059] S3: Construction of a Switching Policy Network and Action Output Based on Deep Reinforcement Learning: Aiming to learn the implicit mapping of "yesterday's pattern—next day's risk—prediction pattern," a Deep Q-Network (DQN) is constructed as the switching policy network. The DQN includes components for processing morphological feature tensors. The convolutional branch, and the branch used to handle calendar priors. The multilayer perceptron branch; outputs action value vectors through the feature fusion layer. and based on Greedy strategy or greedy strategy selection action Output the switching result, where Indicates the target date A point prediction process is adopted. Indicates the target date A probabilistic prediction process is adopted; preferably, the action space supports uniform binary actions across the entire site or differentiated bitmask actions across multiple sites to adapt to different engineering scenarios.

[0060] S4: Reward Function Construction and Offline Training: The core objective is to form a learnable optimization target without embedding specific points / probability predictors, based on the target daily true power curve. Calculate volatility index And based on the volatility distribution of the training set, a volatility threshold is constructed for each site. ;according to With site threshold Constructing point prediction alternative loss Alternative loss for probabilistic prediction Among them, the point prediction substitution loss has a higher penalty slope for the fluctuation exceeding the threshold, while the probability prediction substitution loss has a lower penalty slope for the fluctuation exceeding the threshold and uses a cost term in addition to the probability prediction. The reward is a negative average substitution loss. Used to update DQN parameters; offline training is performed using experience replay and target network mechanisms to obtain the switching strategy after training, and a dual DQN structure is preferred to improve training stability.

[0061] S5: Automatic Policy Output and Prediction Process Decomposition for Validation Set: Aiming to create an engineering strategy file that can directly drive the prediction system, the samples are divided into a training set and a validation set (preferably in a 8:2 ratio) in chronological order. The switching strategy is trained on the training set; the state is constructed daily on the validation set. The trained policy network outputs the switching results and writes them to the policy file; based on the policy file, the power prediction task for the target day is automatically offloaded: when When the point prediction model is invoked, the point prediction results are output. The system can call the probability prediction model to output confidence intervals or scenario results, thereby achieving synergistic optimization of prediction accuracy, robustness, and computing cost.

[0062] Preferably, the sub-step of S1 is as follows:

[0063] S1.1: Reading or capturing data containing... The hourly data in the station power column is matched with timestamps and deduplicated.

[0064] S1.2: Extract power sampling points from 06:00 to 18:00 every day. If there are missing times, resample and fill them in by hour index to make the sequence length of each day fixed at 13.

[0065] S1.3: Perform time-indexed linear interpolation or spline interpolation on missing power points, and use forward padding, backward padding or zero padding on data that is still unrecoverable.

[0066] S1.4: Stack the daily samples in chronological order to form the tensor required for training. , where N is the number of sample days and 13 is the number of sampling points within a day.

[0067] Preferably, the sub-step of S2 is as follows:

[0068] S2.1: For The results were obtained by normalizing the site using the maximum value, quantiles, or scale parameter. To emphasize form while downplaying capacity differences;

[0069] S2.2: Calculation The slope intensity characteristics are obtained by taking the absolute value of the difference between adjacent time points. To characterize mutations and jitters;

[0070] S2.3: Will and Formed by stacking according to channels ;

[0071] S2.4: Construct the target date Among them, the preferred location feature within the year is adopted , The weekday feature uses 7-dimensional one-hot encoding;

[0072] S2.5: Will and Together they constitute a state of no predictive decision-making. .

[0073] Preferably, the processing steps for the action value vector output by the Deep Q-Network (DQN) in S3 include:

[0074] S3.1: The convolutional branch uses at least three two-dimensional convolutional layers. Feature extraction is performed, and a batch normalization layer and a non-linear activation function are set after each convolutional layer to output the image embedding vector.

[0075] S3.2: Prior branches use at least two fully connected layers. Perform feature mapping and output the prior embedding vector;

[0076] S3.3: Concatenate the image embedding vector and the prior embedding vector and input them into the fusion layer to output the action value vector. Where K is the dimension of the action space, and is expressed as... Get the switching action.

[0077] Preferably, the volatility index in S4 The construction includes: calculating the ratio of the average absolute value of adjacent differences to the relative mean of the target daily power sequence, and / or calculating the ratio of the standard deviation of the power sequence to the relative mean, and / or calculating morphological indicators such as peak-to-valley difference, ramp extreme value, and number of multiple peaks, in order to comprehensively reflect the intraday volatility risk.

[0078] Preferably, the threshold in S4 The location is determined by the quantiles of the volatility distribution in the training set, i.e. ,in This is a preset quantile parameter used to distinguish between typical days and fluctuating days.

[0079] Preferably, the substitution loss function in S4 is:

[0080] Point prediction substitution loss

[0081] Probabilistic prediction substitution loss

[0082] in Use a cost term for probabilistic predictions to suppress over-enabling of probabilistic predictions; The preset coefficient is used; the reward is defined as follows: This enables strategy learning for predicting typical daily priority points and predicting fluctuating daily priority probabilities.

[0083] Preferably, the offline training of S4 employs a deep Q-learning algorithm, including: using... A greedy strategy is used to balance exploration and exploitation; an experience replay mechanism is used for small-batch random sampling updates; a stable target Q value is constructed using a target network mechanism or a dual DQN mechanism; the Huber loss is used to minimize the Q value estimation error and the network parameters are updated through the Adam optimizer, thereby obtaining a deployable predictionless switching strategy.

[0084] This invention has the following features and effects:

[0085] 1. It does not rely on future weather information and is highly adaptable.

[0086] By extracting the previous day's morphological features and calendar priors to construct the decision state, the system can operate stably without future meteorological data, effectively solving the switching failure problem in scenarios with missing or unstable meteorological data, and improving the robustness of the system in completely blind scenarios.

[0087] 2. Achieve intelligent adaptive switching to improve prediction accuracy and robustness.

[0088] Based on deep reinforcement learning, a switching strategy network is constructed that can automatically learn the implicit relationship between "yesterday's pattern - next day's risk - prediction mode" and realize the intelligent selection between point prediction and probability prediction. On volatile days, probability prediction is automatically enabled to reduce risk, while point prediction is given priority on typical days to save costs.

[0089] 3. Integrate cost constraints to achieve a balance between economy and accuracy.

[0090] By introducing a cost term for probabilistic prediction into the reward function and jointly optimizing it with a volatility threshold and an alternative loss function, the system can effectively suppress the overuse of probabilistic prediction while ensuring prediction robustness, thereby reducing overall computational and operational costs.

[0091] 4. Decoupled structure, high engineering deployability

[0092] The switching strategy is decoupled from the specific prediction model, eliminating the need to embed it within the predictor. The prediction process is driven by the strategy file, simplifying system integration and maintenance. It supports unified action output across the entire site or differentiated action output across multiple sites, making it easy to deploy and expand in actual photovoltaic power plants.

[0093] 5. Supports differentiated decision-making across multiple sites, flexibly adapting to complex scenarios.

[0094] The action space can be configured as a unified bitmask for the entire station or a station-specific bitmask, enabling differentiated switching based on the fluctuation characteristics of different stations, further improving the overall prediction efficiency and resource utilization efficiency under multi-power station cluster operation conditions.

[0095] Another aspect of the present invention provides a photovoltaic power station-probabilistic joint prediction adaptive switching system based on reinforcement learning, comprising: a computer-readable storage medium and a processor;

[0096] The computer-readable storage medium is used to store executable instructions;

[0097] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the reinforcement learning-based photovoltaic power station-probabilistic joint prediction adaptive switching method described in the first aspect.

[0098] In another aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the reinforcement learning-based photovoltaic power station-probabilistic joint prediction adaptive switching method described in the first aspect.

[0099] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0100] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0101] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0102] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A photovoltaic power station-probabilistic joint prediction adaptive switching method based on reinforcement learning, characterized in that... Includes the following steps: S1. Historical Photovoltaic Power Output Data Acquisition and Intraday Sample Construction: When the photovoltaic power station is in normal operation or in a state awaiting prediction, acquire the historical power output time series data of the photovoltaic power station. The historical power output time series data includes a timestamp and the photovoltaic active power sequence of multiple stations. Reconstruct the historical power output time series data according to a preset intraday time window to form a daily power station output matrix. ; S2. Extraction of morphological features and construction of state variables from the previous day: Extracting the previous day's force matrix. Normalization and differencing are performed to extract characteristic tensors representing the force pattern and wave properties. The feature tensor includes normalized output trajectory features and climbing intensity features; simultaneously, the calendar prior feature vector of the target day d is extracted. The calendar prior feature vector Including the sine and cosine characteristics of the position within the year and the characteristics of the week; and Together they constitute a state of no predictive decision-making. ; S3. Construction and Output of the Switching Policy Network Based on Deep Reinforcement Learning: Construct a Deep Q-Network (DQN) as the switching policy network. The DQN includes components for processing... The convolutional neural network branch and the branch used for processing The multilayer perceptron branch outputs an action value vector at the fusion layer. based on Greedy strategy or greedy strategy selection action And output the binary switching result, where Indicates the target date Using a point prediction model, Indicates the target date A probabilistic prediction model is used; S4. Reward Function Construction and Offline Training: Based on the Actual Output Curve of the Target Day d Calculate volatility index And based on the volatility distribution of the training set, a site threshold is constructed for each site. ;according to With site threshold Calculate the substitution loss function between point prediction cost and probability prediction cost, and take into account the cost of probability prediction to form the reward. Used to update DQN parameters; the DQN is trained offline through experience replay and target network mechanism to obtain the training switching strategy; S5. Automatic strategy output and prediction process separation for validation sets: Generate status for the validation set daily in chronological order. The trained switching strategy is invoked to output the switching results for each target day and write them into the strategy file. Based on the strategy file, the prediction task for the target day is automatically diverted to the point prediction model or the probability prediction model for execution, generating the corresponding point prediction results or probability prediction results.

2. The photovoltaic power station-probabilistic joint prediction adaptive switching method based on reinforcement learning according to claim 1, characterized in that: The sub-steps of S1 are: S1.1: Collect hourly photovoltaic power data by site and align and deduplicate the timestamps; S1.2: Select sample points from 06:00 to 18:00 each day and resample them to keep the number of sample points fixed at 13 per day; S1.3: Perform interpolation and imputation processing on missing points within the day. The interpolation is preferably linear interpolation or spline interpolation based on time index. S1.4: Forming a daily stacked sample tensor , where N is the number of sample days and 13 is the number of sampling points within a day.

3. The photovoltaic power station-probabilistic joint prediction adaptive switching method based on reinforcement learning according to claim 1, characterized in that: The sub-steps of S2 are: S2.1: For the previous day's power matrix Normalize the trajectory features by the maximum value or quantile at each station to obtain the normalized trajectory features. ; S2.2: For the normalized trajectory features Calculate the difference between adjacent time points and take the absolute value to obtain the climbing intensity characteristics. ; S2.3: Will and Formation of morphological feature tensors by stacking channels ; S2.4: Calculate the calendar prior features of the target date The calendar prior features Including positions within the year , And the one-hot coding features of the week; S2.5: Will and Together they constitute a state of no predictive decision-making. .

4. The photovoltaic power station-probabilistic joint prediction adaptive switching method based on reinforcement learning according to claim 1, characterized in that: The processing steps for the action value vector output by the Deep Q-Network (DQN) described in S3 include: S3.1: Convolutional branches use at least three convolutional layers. Feature extraction is performed, and a batch normalization layer and a non-linear activation function are set after each convolutional layer to obtain the image embedding vector. S3.2: Prior branches use at least two fully connected layers. Perform feature mapping to obtain the prior embedding vector; S3.3: Concatenate the image embedding vector with the prior embedding vector and input it into the fusion fully connected layer to output the action value vector. , where K is the dimension of the action space.

5. The photovoltaic power station-probabilistic joint prediction adaptive switching method based on reinforcement learning according to claim 4, characterized in that: The action space of S3.3 is one of the following: (1) A unified action space for the entire site. The corresponding output is {0,1}; (2) Multi-site differentiated action space, The action 'a' encodes the site switching result in the form of a bitmask, and the j-th bit of 'a' indicates whether the j-th site uses point prediction or probability prediction on the target day.

6. The photovoltaic power station-probabilistic joint prediction adaptive switching method based on reinforcement learning according to claim 1, characterized in that: S4 volatility index The construction includes one or a combination of the following: (1) The ratio of the average absolute climbing intensity to its relative mean based on the power difference between adjacent times within the day; (2) Based on the ratio of the standard deviation of the intraday power series to its relative mean; (3) Morphological fluctuation indicators based on intraday peak-to-valley difference, climbing extremes, or the number of multiple peaks: ； in, Indicates the first The first day A daily segment or statistics window, Indicates the switching strategy. For policy network parameters, Represents the expectation operator. This represents the loss function.

7. The photovoltaic power station-probabilistic joint prediction adaptive switching method based on reinforcement learning according to claim 1, characterized in that: Station thresholds in S4 The construction method is as follows: quantile thresholds are calculated for each station of the training set volatility samples. ,in The preset quantile parameter is used to distinguish between typical days and fluctuating days. Used to depict the site Typical fluctuation boundaries; during global handover, the thresholds of each site are aggregated into a site threshold through a weighted average. .

8. The photovoltaic power station-probabilistic joint prediction adaptive switching method based on reinforcement learning according to claim 1, characterized in that: In S4, the reward r(d) is constructed using an alternative loss function, which includes: (1) Point prediction substitution loss ; (2) Probability prediction substitution loss ; in Use cost terms for probabilistic prediction. For preset coefficients, These represent the basic cost of probabilistic prediction and the sensitivity coefficient of probabilistic prediction to out-of-threshold fluctuations, respectively. These represent the base cost of point prediction and the slope of the point prediction over-threshold penalty, respectively. The reward is defined as: ； in, For the first Daily reinforcement learning reward For the number of sites, For site indexing, This means that the corresponding alternative loss function is invoked based on the pattern selected by the site. Site In the The volatility index of the day Representative site The threshold.

9. The photovoltaic power station-probabilistic joint prediction adaptive switching method based on reinforcement learning according to claim 1, characterized in that: S4's offline training employs a deep Q-learning algorithm, including: (1) Adopt Greedy strategy selects actions and gradually decays them. To achieve a balance between exploration and utilization; (2) Use an experience replay mechanism to randomly select a small batch of samples from historical interaction samples for parameter updates; (3) Use a target network mechanism or a dual DQN mechanism to calculate the target Q value to improve training stability; (4) Minimize the Q-value estimation error by using Huber loss or mean squared error loss, and update the network parameters by using Adam optimizer.

10. The photovoltaic power station-probabilistic joint prediction adaptive switching method based on reinforcement learning according to claim 1, characterized in that: The automatic policy output for the validation set in S5 includes: [the output of the validation set's [number]th...] The day before yesterday, the power matrix Build status The switching results are output by the trained DQN and written to the policy file to drive the automatic splitting of subsequent point prediction and probability prediction processes.

Images