A photovoltaic power station power prediction method and system

Abstract

The application relates to the technical field of photovoltaic power generation, and particularly provides a photovoltaic power station power prediction method and system, which comprises the following steps: periodically collecting environment parameters and actual power of a photovoltaic power station, and arranging the collected environment parameters according to the collection time to obtain a parameter sequence; generating predicted power according to the parameter sequence by using a prediction model which is trained based on a meta-learning architecture in advance, wherein the prediction model comprises a main strategy model and multiple sub-strategy models, and the multiple sub-strategy models comprise a short-term power predictor, an ultra-short-term power predictor, a real-time power predictor, a persistence predictor and an average value predictor; calculating the error of the predicted power and the corresponding actual power, and feeding back the error to the prediction model to correct the prediction model. The short-term power prediction model is constructed based on the meta-learning structure, so that short-term power prediction of the photovoltaic power station is realized, and the power prediction precision is improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation technology, and specifically to a method and system for predicting the power output of a photovoltaic power plant. Background Technology

[0002] With the large-scale grid connection of renewable energy power, ensuring the stable and safe operation of the power system is urgently needed. Improving the reliability of photovoltaic power generation systems is a prerequisite for the development of large-scale photovoltaic power generation. Accurate prediction of the short-term output power of photovoltaic power plants can provide advanced guidance for power system dispatch. Summary of the Invention

[0003] In view of the above-mentioned shortcomings of the prior art, the present invention provides a photovoltaic power plant power prediction method and system to solve the above-mentioned technical problems.

[0004] In a first aspect, the present invention provides a method for predicting the power output of a photovoltaic power plant, comprising:

[0005] Regularly collect environmental parameters and actual power of photovoltaic power plants, and sort the collected environmental parameters in chronological order to obtain a parameter sequence.

[0006] The predicted power is generated based on the parameter sequence using a prediction model pre-trained on a meta-learning architecture. The prediction model includes a main policy model and multiple sub-policy models, including a short-term power predictor, an ultra-short-term power predictor, a real-time power predictor, a persistent predictor, and an average value predictor.

[0007] The error between the predicted power and the corresponding actual power is calculated, and the error is fed back to the prediction model to correct the prediction model.

[0008] Furthermore, the training method for the main policy model includes:

[0009] A main strategy objective function is constructed, and the objective function selects a sub-strategy model based on the characteristics of historical power data to obtain the predicted value;

[0010] Based on the given state space dimension of the master policy, the mean square error of the prediction under the selected sub-policy is calculated using the Maldives chain transformation.

[0011] A deep network agent is trained using the gradient stochastic descent algorithm as the main policy objective function. The deep network agent consists of an input layer, an output layer, and multiple hidden layers. The main policy model obtained after training convergence has the minimum prediction mean square error.

[0012] Furthermore, the main policy objective function is constructed, including:

[0013] The objective function of the main policy is constructed based on the principle of selecting a sub-policy that minimizes the mean squared error of the predicted values. Where, without loss of generality, we use 0 to represent the instantaneous decision time; τ = 100ms is the time interval for predicting values; H' represents the main policy; m'≥0 represents the dimension of the state space of the main policy; H' m' (vτ) represents the characteristics of historical power data; H m (vτ) is the power data feature map; n is the time period; a(v) is the data sequence that can be reached within the v-th sampling period; the sub-strategy itself performs prediction based on the feature map.

[0014] Furthermore, based on the given dimensions of the master policy state space, the mean squared error of the prediction under the selected sub-policy is calculated using the Maldives chain transformation, including:

[0015] The optimization of the main policy objective function is expressed as a Maldives chain MDP(S,K,P) k (s,s'),R k (s,s'));

[0016] For a given dimension of the master policy state space, the Maldives chain state space is:

[0017] For the selected sub-policy k, from state A m' (n) is converted to A m' The immediate reward for state (n+1) is: Where R k (s,s') represents the state transition probability;

[0018] After obtaining the state transition probabilities of sub-policy k, state A is... m' (n) The optimization formulation of the principal policy objective function is represented by a Maldives chain; the deep network agent learns the historical power data features H' m' (vτ) and power data feature mapping H m (vτ), the sub-policy performs predictions based on data feature mapping, and then continuously explores the minimum value of the objective function.

[0019] Furthermore, the short-term power predictor, ultra-short-term power predictor, and real-time power predictor are all long short-term memory network predictors trained on a dataset with output power of a single type of photovoltaic system.

[0020] Furthermore, the persistence predictor uses the last observation as the predicted value based on the assumption that the time series is persistent; the average predictor uses the average of the data reached in the last m prediction periods as the predicted value.

[0021] Secondly, the present invention provides a photovoltaic power plant power prediction system, comprising:

[0022] The data acquisition unit is used to periodically collect environmental parameters and actual power of the photovoltaic power station, and sort the collected environmental parameters according to the time of collection to obtain a parameter sequence.

[0023] The power prediction unit is used to generate predicted power based on the parameter sequence using a prediction model pre-trained based on a meta-learning architecture. The prediction model includes a main policy model and multiple sub-policy models, including a short-term power predictor, an ultra-short-term power predictor, a real-time power predictor, a persistent predictor, and an average value predictor.

[0024] The model correction unit is used to calculate the error between the predicted power and the corresponding actual power, and to feed the error back to the prediction model to correct the prediction model.

[0025] Furthermore, the power prediction unit includes:

[0026] The function construction module is used to construct the main strategy objective function, which selects a sub-strategy model based on historical power data features to obtain the predicted value.

[0027] An optimized building module is used to calculate the mean squared error of prediction under the selected sub-policy based on the given main policy state space dimension and the Maldives chain transformation.

[0028] The training execution module is used to train a deep network agent using the gradient stochastic descent algorithm as the main policy objective function. The deep network agent consists of an input layer, an output layer, and multiple hidden layers. The prediction mean square error of the main policy model obtained after training convergence is minimized.

[0029] Furthermore, the function building module is used for:

[0030] The objective function of the main policy is constructed based on the principle of selecting a sub-policy that minimizes the mean squared error of the predicted values. Where, without loss of generality, we use 0 to represent the instantaneous decision time; τ = 100ms is the time interval for predicting values; H' represents the main policy; m'≥0 represents the dimension of the state space of the main policy; H' m' (vτ) represents the characteristics of historical power data; H m (vτ) is the power data feature map; n is the time period; a(v) is the data sequence that can be reached within the v-th sampling period; the sub-strategy itself performs prediction based on the feature map.

[0031] Furthermore, the optimized building module is used for:

[0032] The optimization of the main policy objective function is expressed as a Maldives chain MDP(S,K,P) k(s,s'),R k (s,s'));

[0033] For a given dimension of the master policy state space, the Maldives chain state space is:

[0034] For the selected sub-policy k, from state A m' (n) is converted to A m' The immediate reward for state (n+1) is: Where R k (s,s') represents the state transition probability;

[0035] After obtaining the state transition probabilities of sub-policy k, state A is... m' (n) The optimization formulation of the principal policy objective function is represented by a Maldives chain; the deep network agent learns the historical power data features H' m' (vτ) and power data feature mapping H m (vτ), the sub-policy performs predictions based on data feature mapping, and then continuously explores the minimum value of the objective function.

[0036] The beneficial effect of the present invention is that the photovoltaic power plant power prediction method and system provided by the present invention constructs a short-term power prediction model based on the meta-learning structure, thereby realizing short-term power prediction of photovoltaic power plants and improving the power prediction accuracy.

[0037] Furthermore, the design principle of this invention is reliable, the structure is simple, and it has a very wide range of application prospects. Attached Figure Description

[0038] 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a schematic flowchart of a method according to an embodiment of the present invention.

[0040] Figure 2 This is a schematic flowchart illustrating the meta-learning training of a method according to an embodiment of the present invention.

[0041] Figure 3 This is a schematic architecture diagram of a method according to an embodiment of the present invention. Detailed Implementation

[0042] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0043] Figure 1 This is a schematic flowchart illustrating a method according to an embodiment of the present invention. Wherein, Figure 1 The implementing entity can be a photovoltaic power plant power prediction system.

[0044] like Figure 1 As shown, the method includes:

[0045] Step 110: Periodically collect environmental parameters and actual power of the photovoltaic power station, and sort the collected environmental parameters in chronological order to obtain a parameter sequence;

[0046] Step 120: Generate predicted power based on the parameter sequence using a prediction model pre-trained on a meta-learning architecture. The prediction model includes a main policy model and multiple sub-policy models, including a short-term power predictor, an ultra-short-term power predictor, a real-time power predictor, a persistent predictor, and an average value predictor.

[0047] Step 130: Calculate the error between the predicted power and the corresponding actual power, and feed the error back to the prediction model to correct the prediction model.

[0048] Because photovoltaic (PV) power generation is unstable, using a single prediction model for PV power generation results in poor prediction accuracy under varying PV power generation environments. This invention constructs a prediction model based on a meta-learning structure, which can adapt to various application scenarios of PV power generation, significantly improving short-term prediction accuracy and efficiency.

[0049] For details, please refer to Figure 2 and Figure 3 The photovoltaic power plant power prediction method includes:

[0050] S1. Periodically collect environmental parameters and actual power of the photovoltaic power station, and sort the collected environmental parameters in chronological order to obtain a parameter sequence.

[0051] Environmental sensors are used to periodically collect parameters such as illuminance and temperature, while simultaneously acquiring the operating parameters of the photovoltaic power station. The actual output power is extracted from these operating parameters. The acquisition cycles for environmental parameters and actual power are synchronized. The collected environmental parameters and actual power are saved as parameter sequences and actual power sequences, respectively, according to the acquisition time.

[0052] S2. Generate predicted power based on the parameter sequence using a pre-trained prediction model based on a meta-learning architecture. The prediction model includes a main policy model and multiple sub-policy models, including a short-term power predictor, an ultra-short-term power predictor, a real-time power predictor, a persistent predictor, and an average value predictor.

[0053] Meta-learning, with its core idea of ​​learning how to learn, fully leverages past experience to guide future tasks and is considered key to achieving general artificial intelligence. The meta-learning scheme makes reinforcement learning algorithms robust to adversarial examples and is considered for solving unknown tasks with known distributions in machine learning. The feedback architecture based on the meta-learning scheme selects the learning method based on observed prediction accuracy, allowing the selection of the best predictor for the system at any given time. The meta-learning scheme is designed to consist of a main policy μ∈M and a set of sub-policies k=|K|, such as... Figure 2 As shown. Each sub-policy k∈K is a predictor π trained using photovoltaic power data for a certain period. k This is achieved through a meta-learning scheme. The master policy is responsible for deciding which sub-policy should be used for prediction in the next prediction interval. The basic principle of using a meta-learning scheme to predict non-stationary photovoltaic power plant power is based on its recent application in robust adversarial learning. In this scheme, the master policy is selected from a set of reinforcement learning agents; each agent is trained to solve a specific Markov decision-making (MDP) process, some of which may be perturbed by an adversary. The stochastic photovoltaic output power can be intuitively interpreted as an adversary against individual predictors trained to predict output power with specific characteristics: the adversary presents the predictors with system output power of different characteristics, thus increasing the prediction error. The role of the master policy is to select a good predictor to counter this "adversarial" behavior.

[0054] At any given time point, the master policy aims to select the optimal sub-policy for the next prediction interval. Given that predictors, such as Long Short-Term Memory networks, possess memory, the currently selected sub-policy may influence the accuracy of future sub-policies. Therefore, the goal of the master policy is to find a strategy... To minimize the mean squared error (MSE), i.e.:

[0055]

[0056] Where losses are disregarded, 0 represents the instantaneous decision time, and τ = 100ms is the time interval for predicting values; H' represents the main policy; m'≥0 represents the dimension of the state space of the main policy; H' m' (vτ) represents the characteristics of historical power data; H m (vτ) represents the power data feature map; n is the time period; a(v) is the data sequence that can be reached within the v-th sampling period. It is worth noting that... In other words, the main strategy is based on the historical power data feature H' of m'≥0. m' (vτ) Select a sub-policy This sub-strategy itself is based on the feature mapping H m (vτ) Perform the prediction. Typically, the m' used by the main policy need not be the same as the prediction order m used by the sub-policies.

[0057] The problem of the main strategy can be formulated as a Maldives chain MDP(S,K,P) k (s,s'),R k (s,s')). For a given m', the state space of the MDP is If the action is k, choose action k (i.e., sub-policy). ) and from state A m' (n) is converted to A m' The immediate reward for state (n+1) is:

[0058]

[0059] This is the square of the prediction error. m' is the dimension of the principal policy state space; R is the state transition probability. k The magnitude of (s,s') depends on the photovoltaic system power. After obtaining the state transition probabilities of sub-policy k, state A is... m' (n) The optimization formulation of the principal policy objective function is represented by a Maldives chain; the deep network agent learns the historical power data features H' m' (vτ) and power data feature mapping H m (vτ), the sub-policy performs predictions based on data feature mapping, and then continuously explores the minimum value of the objective function.

[0060] Training the Master Policy: In the case of short-term photovoltaic system power prediction, the state transition probabilities are unknown; therefore, deep reinforcement learning is used to learn the master policy. The master policy selects a predictor and evaluates the reward for each prediction interval. Note that since predictions are performed in real-time at each prediction interval, and the reward values ​​are directly observable, learning can be performed in real-time. Based on this, the master policy can be trained using real-time photovoltaic system output power data and retrained according to actual operating conditions. In the meta-learning scheme, each sub-policy k∈K is a predictor π trained using photovoltaic power data for a certain period. k This is achieved through various methods. Each sub-strategy can be pre-trained or trained in real-time together with the main strategy based on the actual photovoltaic system and the prediction results it provides.

[0061] A complete meta-learning architecture for photovoltaic power plant power prediction, such as Figure 3 As shown:

[0062] A deep Q-network (DQN) agent was trained using stochastic gradient descent as the master policy. The DQN consists of an input layer, an output layer, and multiple hidden layers. Each layer has many neurons. Parameters such as the number of hidden layers, the number of neurons per layer, the learning rate, and the number of learning steps all affect the training model and prediction performance. The stochastic gradient descent algorithm influences the convergence of learning by controlling the weights updated in each batch. Finally, the number of steps is the number of samples used for training, typically millions, achieved by training on a specific dataset over multiple epochs. Let H' be the input feature of the DQN. m' (nτ)=A m' (n).

[0063] The proposed meta-learning scheme is evaluated using five predictor factors as sub-policies. The first three predictors are Long Short-Term Memory (LSTM) network predictors trained on a dataset with output power of a single type of photovoltaic system: the Short-Term Power Predictor, the Ultra-Short-Term Power Predictor, and the Real-Time Power Predictor. The difference between these three predictors lies in the length of their prediction periods.

[0064] Considering that these predictors may perform poorly under various operational environments, two additional predictors are introduced:

[0065] (1) Persistent Model Prediction (PMF)

[0066] PMF assumes the time series is persistent, therefore it uses the latest observation as the forecast. The power forecast is calculated according to the following formula:

[0067] π PMF (n+1)=A(n) (3)

[0068] (2) Average Value Predictor (AVP)

[0069] The AVP output is the average of the data reached over the last m prediction periods. It is derived from a recursive formula:

[0070]

[0071] S3. Calculate the error between the predicted power and the corresponding actual power, and feed the error back to the prediction model to correct the prediction model.

[0072] The meta-learning scheme for photovoltaic (PV) power prediction employs a DQN algorithm with a Q-agent as its main strategy. The PV system output power serves as the data source for the main strategy. The Q-agent continuously learns about the actual operating environment of the PV system, selects appropriate predictors, and receives immediate rewards based on environmental feedback. The five predictors within the sub-strategies learn how to learn based on environmental information and feed their prediction results back to the Q-agent of the main strategy. This iterative update process aims to improve the accuracy of PV system power prediction.

[0073] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the invention should also be covered within the protection scope of the invention. Therefore, the protection scope of the invention should be determined by the scope of the claims.

Claims

1. A method for predicting the power output of a photovoltaic power plant, characterized in that, include: Regularly collect environmental parameters and actual power of photovoltaic power plants, and sort the collected environmental parameters in chronological order to obtain a parameter sequence. The predicted power is generated based on the parameter sequence using a prediction model pre-trained on a meta-learning architecture. The prediction model includes a main policy model and multiple sub-policy models, including a short-term power predictor, an ultra-short-term power predictor, a real-time power predictor, a persistent predictor, and an average value predictor. Calculate the error between the predicted power and the corresponding actual power, and feed the error back to the prediction model to correct the prediction model; The training methods for the main policy model include: A main strategy objective function is constructed, and the objective function selects a sub-strategy model based on the characteristics of historical power data to obtain the predicted value; Based on the given state space dimension of the master policy, the mean square error of the prediction under the selected sub-policy is calculated using the Maldives chain transformation. A deep network agent is trained using the gradient stochastic descent algorithm as the main policy objective function. The deep network agent consists of an input layer, an output layer, and multiple hidden layers. The prediction mean square error of the main policy model obtained after training convergence is minimized. Based on the given dimensions of the master policy state space, the mean squared error of the prediction under the selected sub-policy is calculated using the Maldives chain transformation, including: The optimization of the main policy objective function is expressed as a Maldives chain MDP(S,K,P) k (s,s'),R k (s,s')); For a given dimension of the master policy state space, the Maldives chain state space is: For the selected sub-policy k, from state A m' (n) is converted to A m' The immediate reward for state (n+1) is: Where R k (s,s') represents the state transition probability; After obtaining the state transition probabilities of sub-policy k, state A is... m' (n) The optimization formulation of the principal policy objective function is represented by a Maldives chain; the deep network agent learns the historical power data features H' m' (vτ) and power data feature mapping H m (vτ), the sub-policy performs predictions based on data feature mapping, and then continuously explores the minimum value of the objective function.

2. The method according to claim 1, characterized in that, Construct the main policy objective function, including: The objective function of the main policy is constructed based on the principle of selecting a sub-policy that minimizes the mean squared error of the predicted values. Where 0 represents the instantaneous decision time; τ = 100ms is the time interval for predicting values; H' represents the main policy; m'≥0 represents the dimension of the state space of the main policy; H' m' (vτ) represents the characteristics of historical power data; H m (vτ) is the power data feature map; n is the time period; a(v) is the data sequence that can be reached within the v-th sampling period; the sub-strategy itself performs prediction based on the feature map.

3. The method according to claim 1, characterized in that, The short-term power predictor, ultra-short-term power predictor, and real-time power predictor are all long short-term memory network predictors trained on datasets with output power of a single type of photovoltaic system.

4. The method according to claim 1, characterized in that, The persistence predictor uses the last observation as the predicted value based on the assumption that the time series is persistent; the average predictor uses the average of the data reached in the last m prediction periods as the predicted value.

5. A photovoltaic power plant power prediction system, characterized in that, include: The data acquisition unit is used to periodically collect environmental parameters and actual power of the photovoltaic power station, and sort the collected environmental parameters according to the time of collection to obtain a parameter sequence. The power prediction unit is used to generate predicted power based on the parameter sequence using a prediction model pre-trained based on a meta-learning architecture. The prediction model includes a main policy model and multiple sub-policy models, including a short-term power predictor, an ultra-short-term power predictor, a real-time power predictor, a persistent predictor, and an average value predictor. The model correction unit is used to calculate the error between the predicted power and the corresponding actual power, and to feed the error back to the prediction model to correct the prediction model; The power prediction unit includes: The function construction module is used to construct the main strategy objective function, which selects a sub-strategy model based on historical power data features to obtain the predicted value. An optimized building module is used to calculate the mean squared error of prediction under the selected sub-policy based on the given main policy state space dimension and the Maldives chain transformation. The training execution module is used to train a deep network agent using the gradient stochastic descent algorithm as the main policy objective function. The deep network agent consists of an input layer, an output layer, and multiple hidden layers. The prediction mean square error of the main policy model obtained after training convergence is minimized. The optimized construction module is used for: The optimization of the main policy objective function is expressed as a Maldives chain MDP(S,K,P) k (s,s'),R k (s,s')); For a given dimension of the master policy state space, the Maldives chain state space is: For the selected sub-policy k, from state A m' (n) is converted to A m' The immediate reward for state (n+1) is: Where R k (s,s') represents the state transition probability.

6. The system according to claim 5, characterized in that, The function construction module is used for: The objective function of the main policy is constructed based on the principle of selecting a sub-policy that minimizes the mean squared error of the predicted values. Where 0 represents the instantaneous decision time; τ = 100ms is the time interval for predicting values; H' represents the main policy; m'≥0 represents the dimension of the state space of the main policy; H' m' (vτ) represents the characteristics of historical power data; H m (vτ) is the power data feature map; n is the time period; a(v) is the data sequence that can be reached within the v-th sampling period; the sub-strategy itself performs prediction based on the feature map.

Images