Photovoltaic power generation dynamic simulation method and system based on multi-source data fusion
By using a multi-source data fusion method for dynamic simulation of photovoltaic power generation, and by adjusting the equivalent circuit parameters using a mapping neural network and a time-series residual prediction network, the problem of insufficient simulation accuracy of photovoltaic power generation systems under complex operating conditions is solved, and high-fidelity dynamic simulation is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-12
- Publication Date
- 2026-03-27
AI Technical Summary
Existing photovoltaic power generation simulation methods struggle to accurately detect transient response characteristics during complex processes such as sudden weather changes, resulting in insufficient simulation accuracy. Furthermore, data-driven models lack physical constraints, leading to poor interpretability of simulation results.
By constructing a dynamic simulation method for photovoltaic power generation based on multi-source data fusion, the initial equivalent circuit parameters are solved using a mapping neural network, and the physical model error is accurately predicted by combining a time-series residual prediction network. Furthermore, the physical meaning of the simulation process is ensured by adjusting the equivalent circuit parameters through nonlinear weighting.
It improves the accuracy and reliability of simulation prediction for photovoltaic power generation systems under complex operating conditions, while taking into account both physical interpretability and high-precision prediction.
Smart Images
Figure CN121744870A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of dynamic simulation, and in particular relates to a method and system for dynamic simulation of photovoltaic power generation based on multi-source data fusion. Background Technology
[0002] Photovoltaic (PV) power generation is intermittent and fluctuating, and large-scale grid connection poses a severe challenge to the safe and stable operation of the power system. Establishing accurate simulation models of PV power generation systems to predict power output characteristics under different operating conditions is crucial for power grid planning, dispatch control, and stability analysis. Simulation methods for PV power generation are broadly classified into two categories: physical model-driven and data-driven. Physical model-driven methods are based on the semiconductor physical characteristics of PV cells, representing electrical behavior by constructing equivalent circuit models. This method is clear in its meaning and structure, and exhibits good simulation results under steady-state conditions. However, accurately identifying the equivalent circuit parameters is very difficult, and these parameters change with the external environment. Physical models often use parameters or simplified empirical formulas for correction, making it difficult to accurately detect the transient response characteristics of the PV system under complex processes such as sudden weather changes, resulting in insufficient simulation accuracy. Data-driven methods typically utilize machine learning or deep learning algorithms to learn the nonlinear mapping relationship between environmental factors and output power from massive historical operating data, avoiding the complex physical modeling and parameter identification processes. Data-driven models can reliably uncover deep correlations in the data and exhibit high prediction accuracy on specific datasets. When actual operating conditions exceed the coverage of the training samples, simulation accuracy drops sharply. Data-driven models lack explicit physical constraints, and simulation results may violate the fundamental physical laws of photovoltaic systems, resulting in poor interpretability. Therefore, how to reliably integrate the mechanistic advantages of physical models with the fitting capabilities of data-driven models to achieve complementary advantages and thus perform high-fidelity dynamic simulations of photovoltaic power generation systems under complex and variable operating conditions is a key problem that urgently needs to be solved in the current technological field. Summary of the Invention
[0003] This invention proposes a dynamic simulation method for photovoltaic power generation based on multi-source data fusion, which addresses the problem that existing methods struggle to accurately detect the transient response characteristics of photovoltaic systems during complex processes such as sudden weather changes, leading to insufficient simulation accuracy. The method includes the following steps:
[0004] Historical environmental data, historical power data, and real-time environmental data of the photovoltaic power station to be simulated are obtained; based on the historical environmental data and historical power data, a mapping neural network is constructed, and the initial equivalent circuit parameters of the photovoltaic physical model are obtained by using the network and the real-time environmental data.
[0005] A real-time disturbance factor representing the degree of operating condition disturbance is calculated based on real-time environmental data; a time-series residual prediction network is established and trained, with the rate of change of historical environmental data and historical disturbance factor as training inputs, and the difference between historical power and power calculated by the physical model as the training objective; the rate of change of real-time environmental data and the real-time disturbance factor are input into the trained network to obtain the power residual compensation value.
[0006] A parameter adjustment vector is generated based on the power residual compensation value and the real-time disturbance factor, and the initial equivalent circuit parameters are adjusted using the vector to obtain the instantaneous equivalent circuit parameters; wherein, when the real-time disturbance factor exceeds a first threshold, it is used to perform nonlinear weighted correction on the power residual compensation value, and the parameter adjustment vector is generated based on the corrected compensation value.
[0007] Based on the adjusted instantaneous equivalent circuit parameters and the photovoltaic physical model, the photovoltaic simulation output power is calculated.
[0008] Optionally, the step of constructing a mapping neural network based on historical environmental data and historical power data includes:
[0009] A backpropagation neural network is used as the mapping neural network; the light intensity and temperature from historical environmental data are used as the input layer of the network, and the five equivalent circuit parameters of the photovoltaic physical model, namely the photocurrent, are used. Diode reverse saturation current Series resistor Parallel resistors and diode ideal factor As the output layer, historical environmental data and corresponding historical power data are used to train the network weights and biases by minimizing the root mean square error between the power calculated by the physical model and the historical power, and by using a gradient descent algorithm that combines physical equation constraints.
[0010] Optionally, the step of calculating the real-time disturbance factor representing the degree of operating condition disturbance based on real-time environmental data includes:
[0011] Collect the illumination intensity G(t) and G(t-1) at the current time t and the previous time t-1, as well as the temperature T(t) and T(t-1); calculate the dimensionless real-time perturbation factor k(t) using the following formula:
[0012]
[0013] in, and Let be the weighting coefficient, satisfying ; and For reference light intensity and reference temperature, values under standard test conditions are usually taken.
[0014] Optionally, the establishment and training of the temporal residual prediction network, using the rate of change of historical environmental data and historical perturbation factors as training inputs, and the difference between historical power and power calculated by the physical model as the training objective, includes:
[0015] A long short-term memory network is used as the temporal residual prediction network. The input tensor, consisting of the change rate sequence of historical environmental data from multiple consecutive sampling points and the historical perturbation factor sequence calculated based on the historical environmental data, is used as the training input of the network. The difference between the historical power at the corresponding moment and the power calculated by the physical model is used as the training objective, so that the trained network can predict the power residual compensation value based on the change rate of real-time environmental data and the real-time perturbation factor. The Adam optimizer is used for network training.
[0016] Optionally, the step of calculating the photovoltaic simulation output power based on the adjusted instantaneous equivalent circuit parameters and the photovoltaic physical model includes:
[0017] The photovoltaic physical model is a single-diode model; the calculated photovoltaic simulation output power is obtained by substituting the instantaneous equivalent circuit parameters into the output characteristic equation of the single-diode model.
[0018]
[0019] The maximum power point voltage is obtained by solving the equation using a numerical iteration method. and current And calculate the photovoltaic simulation output power. , where q is the elementary charge, k is the Boltzmann constant, and T is the absolute temperature.
[0020] Optionally, when the real-time disturbance factor exceeds a first threshold, the step of performing a nonlinear weighted correction on the power residual compensation value and generating the parameter adjustment vector based on the corrected compensation value includes:
[0021] The initial power is calculated based on the initial equivalent circuit parameters. and The weights W of the nonlinear weighting function are calculated using the following formula:
[0022]
[0023] in, Let k(t) be the first threshold and k(t) be the real-time disturbance factor. and The preset positive coefficient, The rated power of the photovoltaic power station; the original power residual compensation value. Multiplying by the weight W yields the weighted power residual compensation value. .
[0024] Optionally, the step of generating a parameter adjustment vector based on the power residual compensation value and the real-time disturbance factor, and using the vector to adjust the initial equivalent circuit parameters to obtain the instantaneous equivalent circuit parameters, includes:
[0025] The parameter adjustment vector is the parameter adjustment vector. Based on the power residual compensation value and the real-time disturbance factor, the parameter adjustment vector is generated through a preset mapping relationship; the adjustment vector is then mapped to the initial equivalent circuit parameter vector. Add them together to obtain the instantaneous equivalent circuit parameter vector. .
[0026] Furthermore, this invention also relates to a photovoltaic power generation dynamic simulation system based on multi-source data fusion, comprising the following modules:
[0027] The solution module is used to acquire historical environmental data, historical power data, and real-time environmental data of the photovoltaic power station to be simulated; based on the historical environmental data and historical power data, a mapping neural network is constructed, and the initial equivalent circuit parameters of the photovoltaic physical model are obtained by using the network and the real-time environmental data.
[0028] The input module is used to calculate the real-time disturbance factor representing the degree of operating condition disturbance based on real-time environmental data; to establish and train a time-series residual prediction network, using the rate of change of historical environmental data and historical disturbance factor as training input, and the difference between historical power and power calculated by the physical model as training objective; and to input the rate of change of real-time environmental data and the real-time disturbance factor into the trained network to obtain the power residual compensation value.
[0029] The first calculation module is used to generate a parameter adjustment vector based on the power residual compensation value and the real-time disturbance factor, and to adjust the initial equivalent circuit parameters using the vector to obtain the instantaneous equivalent circuit parameters; wherein, when the real-time disturbance factor exceeds a first threshold, it is used to perform nonlinear weighted correction on the power residual compensation value, and to generate the parameter adjustment vector based on the corrected compensation value.
[0030] The second calculation module is used to calculate the photovoltaic simulation output power based on the adjusted instantaneous equivalent circuit parameters and the photovoltaic physical model.
[0031] Preferably, the step of constructing a mapping neural network based on historical environmental data and historical power data includes:
[0032] A backpropagation neural network is used as the mapping neural network; the light intensity and temperature from historical environmental data are used as the input layer of the network, and the five equivalent circuit parameters of the photovoltaic physical model, namely the photocurrent, are used. Diode reverse saturation current Series resistor Parallel resistors and diode ideal factor As the output layer, historical environmental data and corresponding historical power data are used to train the network weights and biases by minimizing the root mean square error between the power calculated by the physical model and the historical power, and by using a gradient descent algorithm that combines physical equation constraints.
[0033] Preferably, the step of calculating the real-time disturbance factor representing the degree of operating condition disturbance based on real-time environmental data includes:
[0034] Collect the illumination intensity G(t) and G(t-1) at the current time t and the previous time t-1, as well as the temperature T(t) and T(t-1); calculate the dimensionless real-time perturbation factor k(t) using the following formula:
[0035]
[0036] in, and Let be the weighting coefficient, satisfying ; and For reference light intensity and reference temperature, values under standard test conditions are usually taken.
[0037] Preferably, the establishment and training of the temporal residual prediction network, using the rate of change of historical environmental data and historical perturbation factors as training inputs, and the difference between historical power and power calculated by the physical model as the training objective, includes:
[0038] A long short-term memory network is used as the temporal residual prediction network. The input tensor, consisting of the change rate sequence of historical environmental data from multiple consecutive sampling points and the historical perturbation factor sequence calculated based on the historical environmental data, is used as the training input of the network. The difference between the historical power at the corresponding moment and the power calculated by the physical model is used as the training objective, so that the trained network can predict the power residual compensation value based on the change rate of real-time environmental data and the real-time perturbation factor. The Adam optimizer is used for network training.
[0039] Preferably, the step of calculating the photovoltaic simulation output power based on the adjusted instantaneous equivalent circuit parameters and the photovoltaic physical model includes:
[0040] The photovoltaic physical model is a single-diode model; the calculated photovoltaic simulation output power is obtained by substituting the instantaneous equivalent circuit parameters into the output characteristic equation of the single-diode model.
[0041]
[0042] The maximum power point voltage is obtained by solving the equation using a numerical iteration method. and current And calculate the photovoltaic simulation output power. , where q is the elementary charge, k is the Boltzmann constant, and T is the absolute temperature.
[0043] Preferably, the step of performing nonlinear weighted correction on the power residual compensation value when the real-time disturbance factor exceeds a first threshold, and generating the parameter adjustment vector based on the corrected compensation value, includes:
[0044] The initial power is calculated based on the initial equivalent circuit parameters. and The weights W of the nonlinear weighting function are calculated using the following formula:
[0045]
[0046] in, Let k(t) be the first threshold and k(t) be the real-time disturbance factor. and The preset positive coefficient, The rated power of the photovoltaic power station; the original power residual compensation value. Multiplying by the weight W yields the weighted power residual compensation value. .
[0047] Preferably, the step of generating a parameter adjustment vector based on the power residual compensation value and the real-time disturbance factor, and using the vector to adjust the initial equivalent circuit parameters to obtain the instantaneous equivalent circuit parameters, includes:
[0048] The parameter adjustment vector is the parameter adjustment vector. Based on the power residual compensation value and the real-time disturbance factor, the parameter adjustment vector is generated through a preset mapping relationship; the adjustment vector is then mapped to the initial equivalent circuit parameter vector. Add them together to obtain the instantaneous equivalent circuit parameter vector. .
[0049] This invention utilizes a neural network to solve for the initial equivalent circuit parameters of a photovoltaic physical model from historical data. By establishing a time-series residual prediction network, it accurately predicts the calculation errors of the physical model under different operating conditions, and uses the compensation value formed by these errors to adjust the equivalent circuit parameters themselves, rather than simply correcting the power results, thus ensuring the clarity of the physical meaning of the simulation process. An input perturbation factor is used to represent the degree of disturbance of environmental conditions, and combined with a nonlinear weighted adjustment mechanism, the prediction accuracy and reliability of the simulation model under complex operating conditions, such as drastic changes in illumination, are improved. This solves the problem of existing methods struggling to balance physical interpretability and high-precision prediction. Attached Figure Description
[0050] Figure 1 A flowchart of the first embodiment;
[0051] Figure 2 This is a schematic diagram of a mapping network;
[0052] Figure 3 This is a schematic diagram for calculating the real-time disturbance factor. Detailed Implementation
[0053] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.
[0054] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.
[0055] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."
[0056] In the first embodiment, the present invention proposes a dynamic simulation method for photovoltaic power generation based on multi-source data fusion, such as... Figure 1 This includes the following steps:
[0057] S1. Obtain historical environmental data, historical power data, and real-time environmental data of the photovoltaic power station to be simulated; based on the historical environmental data and historical power data, construct a mapping neural network, and use the network and real-time environmental data to solve for the initial equivalent circuit parameters of the photovoltaic physical model;
[0058] Historical environmental data includes historical solar irradiance sequences and ambient temperature sequences sampled at the minute or second level. Historical power data consists of the actual active power output sequences corresponding to the environmental data time points collected by the photovoltaic power plant's SCADA system. Real-time environmental data consists of the solar irradiance and ambient temperature values at the current moment to be simulated.
[0059] Using algorithms such as particle swarm optimization, parameters of the single-diode physical model at each historical time point are identified based on historical environmental and power data. This yields a set of optimal equivalent circuit parameters, including photocurrent, diode reverse saturation current, series resistance, parallel resistance, and diode ideality factor. Environmental data from all historical time points are used as input, and the five identified equivalent circuit parameters are used as output to construct and train a multilayer perceptron network. After training, real-time environmental data is input into the network, and the output becomes the initial equivalent circuit parameters. The photovoltaic physical model uses an equivalent circuit to simulate the internal charge transport mechanism and external current-voltage (IV) characteristics of a photovoltaic cell. It includes a photocurrent source, a parallel diode representing the PN junction characteristics, a parallel resistor representing leakage current, and a series resistor representing internal losses. A single-diode model is preferred.
[0060] In an optional embodiment, constructing a mapping neural network based on historical environmental data and historical power data includes:
[0061] A backpropagation neural network is used as the mapping neural network; the light intensity and temperature from historical environmental data are used as the input layer of the network, and the five equivalent circuit parameters of the photovoltaic physical model, namely the photocurrent, are used. Diode reverse saturation current Series resistor Parallel resistors The ideality factor n of the diode is used as the output layer; using historical environmental data and corresponding historical power data, the root mean square error between the power calculated by the physical model and the historical power is minimized, and the network weights and biases are trained by a gradient descent algorithm that combines physical equation constraints.
[0062] Specifically, the backpropagation (BP) neural network is structured with three layers: an input layer, one or more hidden layers, and an output layer. The input layer has two neurons, receiving light intensity data (e.g., 800 W / m²) and temperature data (e.g., 28°C) from external sensors. The hidden layers can be two layers, each containing 16 neurons, using the Modified Linear Unit (ReLU) as the activation function to enhance the network's nonlinear fitting ability. The output layer has five neurons, corresponding to the five parameters of the photovoltaic physical model. , , , ,n.
[0063] The neural network receives environmental data and propagates forward to output five circuit parameters. These five parameters are then substituted into the nonlinear equations of the photovoltaic physical model. Numerical solutions are used to obtain the voltage and current at the maximum power point, thus yielding the model's calculated power. During training, the loss function is the root mean square error between the calculated power and the actual historical power. The gradient of the loss function with respect to the circuit parameters is calculated using automatic differentiation or implicit function differentiation. This error gradient is then backpropagated to the neural network layers using the chain rule. The gradient descent algorithm iteratively updates the network's weights and biases. The gradient descent algorithm, which incorporates physical equation constraints, embeds known physical laws into the neural network's computational graph or loss function. This ensures that the gradient backpropagation path during model training includes the derivative information of the physical equations, and the parameters learned by the network follow physical mechanisms rather than merely fitting data. For example, in photovoltaic applications, the neural network does not directly output power, but outputs physical parameters such as series resistance. These parameters need to be substituted into the single diode circuit equation to calculate the predicted power. During training, based on the error between the predicted power and the actual power, the chain rule is used to first differentiate the circuit equation, that is, to calculate the gradient of the physical constraints. Then, the gradient is passed back to the neural network to update the weights, ensuring that the resistance value output by the network can both enable the equation to solve for the correct power and conform to the physical constraints of the circuit model.
[0064] S2, calculate the real-time disturbance factor representing the degree of operating condition disturbance based on real-time environmental data; establish and train a time-series residual prediction network, using the rate of change of historical environmental data and historical disturbance factor as training inputs, and the difference between historical power and power calculated by the physical model as the training objective; input the rate of change of real-time environmental data and the real-time disturbance factor into the trained network to obtain the power residual compensation value.
[0065] Specifically, the real-time disturbance factor is obtained by calculating the absolute value of the rate of change of the irradiance between the current time and the previous time. For example, if the sampling time interval is Δt and the current irradiance is... The light intensity at the previous moment was Then the real-time disturbance factor equal and The absolute value of the difference is divided by Δt. The temporal residual prediction network preferably employs a Long Short-Term Memory (LSTM) network, aiming to predict the error between the physical model and the actual power based on the dynamic changes in environmental data. During training, the network uses a sequence of historical environmental data change rates (e.g., historical irradiance change rate, ambient temperature change rate) and a sequence of historical perturbation factors calculated from these data as input features. Simultaneously, the residual sequence, representing the difference between the historical actual power and the power calculated by the physical model at the corresponding moment, is used as the output label for training. After training, in the real-time simulation phase, the current real-time environmental data change rate and real-time perturbation factors are input into the network to obtain a prediction of the model calculation error, i.e., the power residual compensation value. In another embodiment, the prediction network structure is as follows... Figure 2 It includes an input layer, a hidden layer, and an output layer.
[0066] In an optional embodiment, the step of calculating the real-time disturbance factor representing the degree of operating condition disturbance based on real-time environmental data includes:
[0067] Collect the illumination intensity G(t) and G(t-1) at the current time t and the previous time t-1, as well as the temperature T(t) and T(t-1); calculate the dimensionless real-time perturbation factor k(t) using the following formula:
[0068]
[0069] in, and Let be the weighting coefficient, satisfying ; and For reference light intensity and reference temperature, values under standard test conditions are usually taken.
[0070] Assuming data is sampled every minute, , One minute prior, at time t-1, the collected light intensity G(t-1) was 900 W / m², and the temperature T(t-1) was 30℃. At the current time t, due to cloud cover, the collected light intensity G(t) has decreased to 600 W / m², and the temperature T(t) has slightly decreased to 29.5℃. This data will be sent to the calculation module.
[0071] The absolute changes in light intensity and temperature were calculated to be 300 W / m² and 0.5 °C, respectively. Normalization was performed using reference values: the normalized change in light intensity was 0.3, and the normalized change in temperature was 0.02. Preset weighting coefficients were applied, for example, assigning a greater weight to changes in light intensity. Temperature weighting The calculated real-time disturbance factor k(t) is 0.244. The dimensionless value of 0.244 represents the severity of the disturbance under the current operating condition. Figure 3 .
[0072] In an optional embodiment, the establishment and training of the temporal residual prediction network, using the rate of change of historical environmental data and historical perturbation factors as training inputs, and the difference between historical power and power calculated by the physical model as the training objective, includes:
[0073] A long short-term memory network is used as the temporal residual prediction network. The input tensor, consisting of the change rate sequence of historical environmental data from multiple consecutive sampling points and the historical perturbation factor sequence calculated based on the historical environmental data, is used as the training input of the network. The difference between the historical power at the corresponding moment and the power calculated by the physical model is used as the training objective, so that the trained network can predict the power residual compensation value based on the change rate of real-time environmental data and the real-time perturbation factor. The Adam optimizer is used for network training.
[0074] The core architecture of the temporal residual prediction network is a Long Short-Term Memory (LSTM) network, planned to include one input layer, two stacked LSTM layers, and a fully connected output layer. Each LSTM layer contains 64 memory units, enabling reliable detection of long-term dependencies in time-series data. The input layer receives a tensor of a specific shape, composed of data from a consecutive time window, such as 10 sampling points. For each sampling point, the feature vector contains three values: the normalized rate of change of light intensity, the normalized rate of change of temperature, and the historical perturbation factor at that moment. Therefore, the input tensor dimension of a training sample is 10×3.
[0075] When training the network, a training dataset is constructed. From a large amount of historical data, for each time point t, a 10×3 input tensor as described above is constructed, with the time range from t-9 to t. The training target or label corresponding to this input tensor is the power residual at time t, that is, the difference between the actual recorded historical power and the power calculated by the physical model using the output parameters of the BP network. A large number of input tensors and target residuals are fed into the network, which outputs a predicted residual value. The mean squared error loss is calculated by comparing the predicted residual with the true residual. The Adam optimizer is used to automatically adjust the learning rate based on the loss value and backpropagates to update the network weights. The training is iterated repeatedly until the network model's error on the validation set no longer decreases.
[0076] S3, Generate a parameter adjustment vector based on the power residual compensation value and the real-time disturbance factor, and use the vector to adjust the initial equivalent circuit parameters to obtain the instantaneous equivalent circuit parameters; wherein, when the real-time disturbance factor exceeds a first threshold, it is used to perform nonlinear weighted correction on the power residual compensation value, and generate the parameter adjustment vector based on the corrected compensation value.
[0077] Specifically, the parameter adjustment vector is a vector containing five adjustment coefficients, each corresponding to one of the five initial equivalent circuit parameters. This adjustment vector is generated by a small fully connected network based on the power residual compensation value and the real-time disturbance factor. When the real-time disturbance factor is less than or equal to a first threshold, such as 0.15, the adjustment vector is directly generated using the method described above. When the real-time disturbance factor exceeds this threshold, a nonlinear weighting function, such as an sigmoid function, is first constructed. The growth rate of this function is proportional to the magnitude of the real-time disturbance factor. A weight greater than 1 is calculated using this function, and this weight is multiplied by the original power residual compensation value output by the LSTM network in step S2 to obtain the weighted compensation value. Then, the parameter adjustment vector is generated using this weighted compensation value and the real-time disturbance factor through the small fully connected network. Finally, the instantaneous equivalent circuit parameters are obtained by element-wise summing the initial equivalent circuit parameter vector and the parameter adjustment vector.
[0078] In an optional embodiment, the step of performing nonlinear weighted correction on the power residual compensation value when the real-time disturbance factor exceeds a first threshold, and generating the parameter adjustment vector based on the corrected compensation value, includes:
[0079] The initial power is calculated based on the initial equivalent circuit parameters. and The weights W of the nonlinear weighting function are calculated using the following formula:
[0080]
[0081] in, Let k(t) be the first threshold and k(t) be the real-time disturbance factor. and The preset positive coefficient, The rated power of the photovoltaic power station; the original power residual compensation value. Multiplying by the weight W yields the weighted power residual compensation value. .
[0082] The real-time calculated disturbance factor k(t) is compared with a preset threshold. Comparison, assumption The value is 0.15. If the calculated k(t) at a certain moment is 0.244, since it is greater than 0.15, the nonlinear weighting process is triggered. A BP network is invoked, and using the environmental data from the current time t and the previous time t-1, two sets of initial equivalent circuit parameters are calculated respectively. The initial power is then calculated using a single diode model. and For example, to obtain , .
[0083] Assuming the rated power of the photovoltaic power station Preset coefficients and The power variation term is 0.25. Substituting this into the formula to calculate the exponent part, we get 0.563. Therefore, the nonlinear weight W = 2.756. The original power residual compensation value predicted by the LSTM network, for example... Multiply this by the weight to obtain the weighted compensation value. This weighted value will be used for subsequent fine-tuning of parameters to cope with drastic changes in operating conditions.
[0084] In an optional embodiment, the step of generating a parameter adjustment vector based on the power residual compensation value and the real-time disturbance factor, and using the vector to adjust the initial equivalent circuit parameters to obtain the instantaneous equivalent circuit parameters, includes:
[0085] The parameter adjustment vector is the parameter adjustment vector. Based on the power residual compensation value and the real-time disturbance factor, the parameter adjustment vector is generated through a preset mapping relationship; the adjustment vector is then mapped to the initial equivalent circuit parameter vector. Add them together to obtain the instantaneous equivalent circuit parameter vector. .
[0086] Specifically, this preset mapping relationship is a key module whose function is to input two scalar values, namely the power residual compensation value. The real-time disturbance factor k(t) is converted into a five-dimensional parameter adjustment vector. This mapping can be a small feedforward neural network or a rule-based lookup table based on physical properties. For example, the rule could be set such that when the power residual... When the value is positive, it indicates that the photocurrent is underestimated, therefore It should be positive, and its size is the same as... Positive correlation; if the disturbance factor k(t) is large, then the series resistance adjustment amount It should also be increased accordingly.
[0087] Assuming the current weighted power residual compensation value is The real-time disturbance factor k(t) = 0.244. The parameter adjustment vector calculated using the preset mapping relationship is... Optionally, the preset mapping relationship is a pre-constructed sensitivity coefficient matrix of power to each circuit parameter, and the power residual compensation value is allocated to the correction amount of each circuit parameter according to the sensitivity ratio. The initial equivalent circuit parameter vector given by the BP neural network is... The final step is to perform vector addition, adding the initial parameter vector and the adjustment vector element by element to obtain the instantaneous equivalent circuit parameter vector. This set of finely tuned instantaneous parameters will be used to calculate the most accurate output power at the current moment.
[0088] In another optional embodiment, the step of generating a parameter adjustment vector based on the power residual compensation value and the real-time disturbance factor, and using the vector to adjust the initial equivalent circuit parameters to obtain the instantaneous equivalent circuit parameters, includes:
[0089] Based on the initial equivalent circuit parameters Using real-time environmental data, the partial derivatives of the photovoltaic physical model's output power with respect to five equivalent circuit parameters are calculated to form a parameter sensitivity vector. ; Calculate the basic adjustment step size based on the power residual compensation value ΔP. The dynamic adjustment step size is obtained by scaling the base adjustment step size using the real-time disturbance factor k(t) through a preset gain function. Multiply the dynamic adjustment step size by the parameter sensitivity vector to generate the parameter adjustment vector. The adjustment vector is compared with the initial equivalent circuit parameter vector. Add them together to obtain the instantaneous equivalent circuit parameter vector. .
[0090] For example, at a certain moment, the photovoltaic physical model is based on initial parameters. The calculated power is 980W, while the timing residual prediction network (LSTM) in step S2 predicts a calculation error of +20W in the physical model under the current operating condition. Therefore, the power residual ΔP that needs to be compensated is +20W. Simultaneously, due to drastic changes in illumination, the calculated real-time disturbance factor k(t) is 0.9. The sensitivity vector S of the power to five circuit parameters under the current state is calculated, for example... This indicates that the photocurrent increased by 1A at this time This can result in a power increase of approximately 50W, while increasing the series resistance by 1Ω. This would cause a power drop of approximately 30W. A basic adjustment step size η is calculated based on ΔP and S. Then, using a high disturbance factor k(t) = 0.9, η is amplified by a factor of 1.9 through a gain function such as 1 + k(t) to obtain a larger dynamic adjustment step size. This is to enable a rapid response to change. Multiplying this by the sensitivity vector S generates a specific parameter adjustment vector Δθ, whose components are distributed according to the proportions of S. For example... This will result in the largest positive adjustment, and This will result in the second largest negative adjustment. Add Δθ to the initial parameters. The instantaneous equivalent circuit parameters that reflect the current actual operating conditions are obtained. This allows the model output to approach 1000W.
[0091] S4. Based on the adjusted instantaneous equivalent circuit parameters and the photovoltaic physical model, the photovoltaic simulation output power is calculated.
[0092] Specifically, the adjusted five instantaneous equivalent circuit parameters, along with real-time environmental data, are substituted into the current-voltage characteristic equation of the photovoltaic single-diode model. Numerical iterative methods, such as Newton-Raphson methods, are used to solve for the maximum power point under the current operating conditions. The power value corresponding to this maximum power point is the photovoltaic simulation output power.
[0093] In an optional embodiment, the step of calculating the photovoltaic simulation output power based on the adjusted instantaneous equivalent circuit parameters and the photovoltaic physical model includes:
[0094] The photovoltaic physical model is a single-diode model; the calculated photovoltaic simulation output power is obtained by substituting the instantaneous equivalent circuit parameters into the output characteristic equation of the single-diode model.
[0095]
[0096] The maximum power point voltage is obtained by solving the equation using a numerical iteration method. and current And calculate the photovoltaic simulation output power. , where q is the elementary charge, k is the Boltzmann constant, T is the absolute temperature, and I and V are the output current and output voltage of the photovoltaic cell, respectively.
[0097] The instantaneous parameter obtained at a certain moment is the photocurrent. A, Reverse saturation current A, Series resistance Parallel resistors With an ideality factor n = 1.21 and a current absolute temperature T of 301.15 K, substituting these values along with the physical constants q and k into the equation yields an implicit equation concerning the output voltage V and the output current I.
[0098] Since this equation cannot be solved analytically directly, a numerical iteration method is used to find the maximum power point. A commonly used method is the voltage scan method. This method sets a voltage scan range, for example, from 0V to the estimated open-circuit voltage of the photovoltaic module, with a step size of 0.1V. For each voltage value... The implicit equations above are solved using numerical solvers such as Newton's iteration method to obtain the corresponding current values. In this way, a series of voltage and current data points can be obtained, which constitute the IV characteristic curve of the photovoltaic module. For each data point, the power is calculated. This yields the PV characteristic curve. The process iterates through all power values. Find the maximum value among them; this maximum value is the photovoltaic simulation output power. The corresponding voltage and current are and .
[0099] In a second embodiment, the present invention also provides a photovoltaic power generation dynamic simulation system based on multi-source data fusion, comprising the following modules:
[0100] The solution module is used to acquire historical environmental data, historical power data, and real-time environmental data of the photovoltaic power station to be simulated; based on the historical environmental data and historical power data, a mapping neural network is constructed, and the initial equivalent circuit parameters of the photovoltaic physical model are obtained by using the network and the real-time environmental data.
[0101] The input module is used to calculate the real-time disturbance factor representing the degree of operating condition disturbance based on real-time environmental data; to establish and train a time-series residual prediction network, using the rate of change of historical environmental data and historical disturbance factor as training input, and the difference between historical power and power calculated by the physical model as training objective; and to input the rate of change of real-time environmental data and the real-time disturbance factor into the trained network to obtain the power residual compensation value.
[0102] The first calculation module is used to generate a parameter adjustment vector based on the power residual compensation value and the real-time disturbance factor, and to adjust the initial equivalent circuit parameters using the vector to obtain the instantaneous equivalent circuit parameters; wherein, when the real-time disturbance factor exceeds a first threshold, it is used to perform nonlinear weighted correction on the power residual compensation value, and to generate the parameter adjustment vector based on the corrected compensation value.
[0103] The second calculation module is used to calculate the photovoltaic simulation output power based on the adjusted instantaneous equivalent circuit parameters and the photovoltaic physical model.
[0104] 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, compact disc read-only memory (CD-ROM), optical storage, etc.) containing computer-usable program code.
[0105] 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.
[0106] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.
Claims
1. A photovoltaic power generation dynamic simulation method based on multi-source data fusion, characterized in that, The method comprises the following steps: acquiring historical environment data, historical power data and real-time environment data of a photovoltaic power station to be simulated; constructing a mapping neural network based on the historical environment data and the historical power data, and using the network and the real-time environment data to obtain initial equivalent circuit parameters of a photovoltaic physical model; calculating a real-time disturbance factor representing the disturbance degree of the working condition according to the real-time environment data; establishing and training a time-series residual prediction network, taking the change rate of the historical environment data and the historical disturbance factor as the training input, and taking the difference between the historical power and the calculated power of the physical model as the training target; inputting the change rate of the real-time environment data and the real-time disturbance factor into the trained network to obtain a power residual compensation value; generating a parameter adjustment vector based on the power residual compensation value and the real-time disturbance factor, and adjusting the initial equivalent circuit parameters using the vector to obtain instantaneous equivalent circuit parameters; when the real-time disturbance factor exceeds a first threshold, the power residual compensation value is nonlinearly weighted and corrected, and the parameter adjustment vector is generated based on the corrected compensation value; calculating the photovoltaic simulation output power based on the adjusted instantaneous equivalent circuit parameters and the photovoltaic physical model.
2. The method of claim 1, wherein, The method comprises the following steps: adopting back propagation neural network as the mapping neural network; taking the light intensity and temperature in the historical environment data as the input layer of the network, taking five equivalent circuit parameters of the photovoltaic physical model, i.e. , diode reverse saturation current , series resistance , parallel resistance and diode ideality factor n as the output layer; using the historical environment data and the corresponding historical power data, the network weight and bias are trained by adopting the gradient descent algorithm combined with the constraint of the physical equation by minimizing the root mean square error of the power calculated by the physical model and the historical power.
3. The method of claim 2, wherein, The method comprises the following steps: Collecting the light intensity G(t) and G(t-1) and the temperature T(t) and T(t-1) at the current time t and the previous time t-1; calculating the dimensionless real-time disturbance factor k(t) by the following formula: wherein, and are weight coefficients satisfying ; and are reference light intensity and reference temperature.
4. The method of claim 3, wherein, The method comprises the following steps: Using a long short-term memory network as the time-series residual prediction network; taking an input tensor composed of the change rate sequence of the historical environment data of a plurality of consecutive sampling points and the historical disturbance factor sequence calculated based on the historical environment data as the training input of the network, and taking the difference between the historical power and the calculated power of the physical model at the corresponding time as the training target, so that the trained network can predict the power residual compensation value according to the change rate of the real-time environment data and the real-time disturbance factor; using an Adam optimizer to train the network.
5. The method of claim 4, wherein, The method comprises the following steps: The photovoltaic physical model is a single-diode model; the method comprises the following steps: where the maximum power point voltage is obtained by solving the equation with a numerical iteration method and current and the photovoltaic simulation output power is calculated where q is the elementary charge, k is the Boltzmann constant, T is the absolute temperature, I and V are the output current and output voltage, respectively.
6. The method according to claim 1 or 3, characterized in that, When the real-time disturbance factor exceeds the first threshold, the power residual compensation value is nonlinearly weighted and corrected, and the parameter adjustment vector is generated based on the corrected compensation value. calculating an initial power from the initial equivalent circuit parameters and ; the weight W of the non-linear weighting function is calculated by wherein, is a first threshold value, k(t) is a real-time disturbance factor, and is a preset positive coefficient, is a rated power of the photovoltaic power station; the original power residual compensation value is multiplied by the weight W to obtain a weighted power residual compensation value.
7. The method of claim 2, wherein, The method comprises the following steps: The parameter adjustment vector is a parameter adjustment vector ; based on the power residual compensation value and the real-time disturbance factor, generating the parameter adjustment vector through a preset mapping relationship; adding the adjustment vector and the initial equivalent circuit parameter vector to obtain an instantaneous equivalent circuit parameter vector .
8. A photovoltaic power generation dynamic simulation system based on multi-source data fusion, characterized in that, The method comprises the following modules: The solving module is configured to acquire historical environment data, historical power data and real-time environment data of a photovoltaic power station to be simulated; based on the historical environment data and the historical power data, a mapping neural network is constructed, and based on the network and the real-time environment data, initial equivalent circuit parameters of a photovoltaic physical model are solved; The input module is configured to calculate a real-time disturbance factor representing a disturbance degree of a working condition according to the real-time environment data; a time-series residual prediction network is established and trained, with a change rate of the historical environment data and a historical disturbance factor as training inputs, and with a difference between historical power and a power calculated by the physical model as a training target; The real-time environment data change rate and the real-time disturbance factor are input into the trained network to obtain a power residual compensation value; The first calculation module is configured to generate a parameter adjustment vector based on the power residual compensation value and the real-time disturbance factor, and adjust the initial equivalent circuit parameters by using the vector to obtain instantaneous equivalent circuit parameters; when the real-time disturbance factor exceeds a first threshold, the power residual compensation value is nonlinearly weighted and corrected, and the parameter adjustment vector is generated based on the corrected compensation value; The second calculation module is configured to calculate a photovoltaic simulation output power based on the adjusted instantaneous equivalent circuit parameters and the photovoltaic physical model.
9. The system of claim 8, wherein, The mapping neural network is constructed based on the historical environment data and the historical power data, including: adopting back propagation neural network as the mapping neural network; taking the light intensity and temperature in the historical environment data as the input layer of the network, taking five equivalent circuit parameters of the photovoltaic physical model, i.e. , diode reverse saturation current , series resistance , parallel resistance and diode ideality factor n as the output layer; using the historical environment data and the corresponding historical power data, the network weight and bias are trained by adopting the gradient descent algorithm combined with the constraint of the physical equation by minimizing the root mean square error of the power calculated by the physical model and the historical power.
10. The system of claim 9, wherein, The real-time disturbance factor representing the disturbance degree of the working condition is calculated according to the real-time environment data, including: The illumination intensity G(t) and G(t-1) and the temperature T(t) and T(t-1) at the current time t and the previous time t-1 are collected; the real-time disturbance factor k(t) is calculated by the following formula: wherein and are weight coefficients, satisfying ; and are reference light intensity and reference temperature, usually taking values under standard test conditions.