An ultra-short-term photovoltaic power prediction method and system based on thermal state perception

Abstract

The application belongs to the technical field of photovoltaic power generation power prediction, and discloses a kind of ultra-short-term photovoltaic power prediction method and system based on thermal state perception, the method includes: obtaining the multi-source heterogeneous measurement data of photovoltaic power station to construct continuous time input path function;And input to the physical neural differential model trained in advance, calculate the continuous time temperature trajectory and heat flux trajectory, and in each prediction interval, extract thermal transient feature, and construct thermal transient feature vector;Energy balance consistency screening is carried out to each historical interval, and a candidate historical interval set is constructed;The cross-attention calculation and fusion processing of thermal similarity bias are carried out to the candidate historical interval set, and the initial power prediction result is obtained;And substitute dynamic energy balance equation to carry out energy balance back substitution check correction, output final photovoltaic power prediction result.The application realizes high-precision continuous time reconstruction, thereby improving the robustness in photovoltaic power ramp and drop scene.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, and in particular to an ultra-short-term photovoltaic power prediction method and system based on thermal state sensing. Background Technology

[0002] With the transformation of the global energy structure, the proportion of photovoltaic power generation in the power system continues to grow. However, the output power of photovoltaic power is significantly affected by the irradiance fluctuations caused by clouds, and passing clouds can cause exceedances of rated capacity within a timescale of seconds. Power fluctuations pose a challenge to grid frequency regulation and voltage stability. To address this challenge, it is necessary to achieve high-precision ultra-short-term photovoltaic power prediction.

[0003] Existing photovoltaic power prediction technologies have the following shortcomings:

[0004] (1) Steady-state thermal assumption leads to systematic bias. Existing physical and hybrid methods typically employ a steady-state thermal assumption, neglecting the thermal inertia effect of photovoltaic modules. Since the thermal time constant of crystalline silicon modules is 5 to 15 minutes, the module temperature cannot respond instantaneously to changes in irradiance. The steady-state assumption produces systematic temperature estimation bias when clouds pass rapidly, thus affecting the accuracy of power prediction. In addition, in actual operation, the measurement data of ambient temperature, irradiance, and wind speed have different sampling frequencies, making it difficult to directly and uniformly substitute them into the continuous-time energy balance equation.

[0005] (2) Discrete sampling leads to the loss of high-frequency dynamic information. Existing prediction methods mostly use discrete sampling data with intervals of 1 minute or longer as input. High-frequency thermal dynamic features such as the sub-minute temperature change rate during cloud passage are filtered out or aliased during the discrete sampling process, resulting in insufficient sensitivity of the model to rapid power fluctuations.

[0006] (3) The attention mechanism lacks physical guidance. Existing attention-based prediction models assign attention weights based on temporal proximity, without considering the similarity of thermophysical states. When weather conditions change rapidly, historical samples that are temporally adjacent may be in different thermal states, while historical samples that are far apart in time but have similar thermal states cannot be effectively retrieved and utilized.

[0007] Therefore, how to provide an ultra-short-term photovoltaic power prediction method and system based on thermal state perception is an urgent problem to be solved. Summary of the Invention

[0008] This invention provides an ultra-short-term photovoltaic power prediction method based on thermal state sensing to solve the problems mentioned above in the prior art.

[0009] According to a first aspect of the present invention, an ultra-short-term photovoltaic power prediction method based on thermal state sensing is provided.

[0010] In one embodiment, the ultra-short-term photovoltaic power prediction method based on thermal state sensing includes:

[0011] Acquire multi-source heterogeneous measurement data from photovoltaic power plants, and perform continuous-time interpolation on the multi-source heterogeneous measurement data to construct a continuous-time input path function;

[0012] The continuous-time input path function is input into a pre-trained physical neural differential model to calculate the continuous-time temperature trajectory and heat flux trajectory of the photovoltaic module; wherein, the physical neural differential model includes a dynamic energy balance equation, a neural network model set, and a dual-scale branch;

[0013] Within each prediction interval, thermal transient features are extracted from the continuous-time temperature trajectory and heat flux trajectory to construct a thermal transient feature vector;

[0014] Based on the thermal transient feature vector, the net thermal driving direction parameters and energy closure residuals for each prediction interval and each historical interval are calculated respectively. Combined with the preset energy balance residual threshold, the energy balance consistency screening process is carried out for each historical interval to construct a candidate historical interval set.

[0015] Using a dual-scale branch, cross-attention calculation with thermal similarity bias is performed on the candidate historical interval set, and adaptive gating fusion is performed based on temperature change rate features to obtain the initial power prediction results;

[0016] The initial power prediction results are substituted into the dynamic energy balance equation for energy balance back-substitution verification and correction to obtain the final photovoltaic power prediction results.

[0017] According to a second aspect of the present invention, an ultra-short-term photovoltaic power prediction system based on thermal state sensing is provided.

[0018] In one embodiment, the ultra-short-term photovoltaic power prediction system based on thermal state sensing includes:

[0019] The continuous path construction module is used to acquire multi-source heterogeneous measurement data from photovoltaic power plants and perform continuous-time interpolation processing on the multi-source heterogeneous measurement data to construct a continuous-time input path function.

[0020] The physical information thermodynamics calculation module is used to input the continuous-time input path function into the pre-trained physical neural differential model to calculate the continuous-time temperature trajectory and heat flux trajectory of the photovoltaic module; wherein, the physical neural differential model includes a dynamic energy balance equation, a neural network model set, and a dual-scale branch;

[0021] The thermal transient feature extraction module is used to extract thermal transient features from the continuous-time temperature trajectory and heat flux trajectory within each prediction interval, and construct a thermal transient feature vector.

[0022] The energy balance screening module is used to calculate the net thermal driving direction parameters and energy closure residuals for each prediction interval and each historical interval based on the thermal transient feature vector, and to perform energy balance consistency screening for each historical interval in combination with the preset energy balance residual threshold, thereby constructing a set of candidate historical intervals.

[0023] The thermal state perception and prediction module is used to perform cross-attention calculation and fusion processing with thermal similarity bias on the candidate historical interval set using dual-scale branching to obtain the initial power prediction result;

[0024] The energy balance back-substitution verification and correction module is used to substitute the initial power prediction results into the dynamic energy balance equation for energy balance back-substitution verification and correction, so as to obtain the final photovoltaic power prediction results.

[0025] According to a third aspect of the present invention, a computer device is provided.

[0026] In one embodiment, the computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0027] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0028] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.

[0029] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0030] 1. By embedding the energy balance equation into the physical neural differential model and fusing input signals with different sampling frequencies through adaptive step-size numerical integration, the systematic temperature estimation bias introduced by the steady-state thermal assumption during rapid cloud passage is overcome, and high-precision continuous-time reconstruction of component temperature is achieved.

[0031] 2. By constructing a thermal transient feature vector that includes the rate of temperature change, average convective heat loss, irradiation-temperature hysteresis characterization quantity, and equivalent thermal time constant, it is possible to distinguish working conditions with the same instantaneous temperature but different heating, cooling and thermal inertia states, thus avoiding the retrieval bias caused by relying solely on the instantaneous thermal state.

[0032] 3. Before cross-attention calculation, energy balance consistency screening is introduced. First, historical intervals with inconsistent energy balance are eliminated based on the net heat driving direction and energy balance residual. Then, cross-attention calculation with thermal similarity bias is performed within the candidate historical interval set, thereby improving the physical reliability of searching for historical intervals with similar thermal responses under complex weather conditions.

[0033] 4. After the initial power prediction result is output by cross-attention, an energy balance back-substitution verification correction is performed to ensure that the final output predicted power and the future temperature trajectory meet the closed-loop consistency constraint, thereby improving the robustness in photovoltaic power ramp-up and ramp-down scenarios.

[0034] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0035] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0036] Figure 1 This is a flowchart illustrating an ultra-short-term photovoltaic power prediction method based on thermal state sensing, according to an exemplary embodiment.

[0037] Figure 2 This is a schematic diagram of an ultra-short-term photovoltaic power prediction system based on thermal state sensing, according to an exemplary embodiment.

[0038] Figure 3 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment;

[0039] Figure 4 This is a flowchart illustrating the calculation of continuous-time temperature trajectory and heat flux trajectory in an ultra-short-term photovoltaic power prediction method based on thermal state perception, according to an exemplary embodiment.

[0040] Figure 5 This is a flowchart illustrating the calculation of the final photovoltaic power prediction result in an ultra-short-term photovoltaic power prediction method based on thermal state perception, according to an exemplary embodiment.

[0041] Figure 6 This is a flowchart illustrating the training process of a physical neural differential model in an ultra-short-term photovoltaic power prediction method based on thermal state perception, according to an exemplary embodiment. Detailed Implementation

[0042] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some portions and features of certain embodiments may be included in or replace portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents thereof. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0043] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0044] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0045] Figure 1 An embodiment of an ultra-short-term photovoltaic power prediction method based on thermal state sensing according to the present invention is shown.

[0046] In this optional embodiment, the ultra-short-term photovoltaic power prediction method based on thermal state sensing includes:

[0047] S101. Obtain multi-source heterogeneous measurement data of photovoltaic power station, and perform continuous-time interpolation processing on multi-source heterogeneous measurement data to construct continuous-time input path function;

[0048] S102. Input the continuous-time input path function into the pre-trained physical neural differential model to calculate the continuous-time temperature trajectory and heat flux trajectory of the photovoltaic module; wherein, the physical neural differential model includes a dynamic energy balance equation, a neural network model set, and a dual-scale branch;

[0049] S103. Within each prediction interval, extract thermal transient features from the continuous-time temperature trajectory and heat flux trajectory, and construct a thermal transient feature vector.

[0050] S104. Based on the thermal transient feature vector, calculate the net thermal driving direction parameters and energy closure residuals for each prediction interval and each historical interval, and combine them with the preset energy balance residual threshold to perform energy balance consistency screening for each historical interval and construct a candidate historical interval set.

[0051] S105. Using the dual-scale branch training, perform cross-attention calculation and fusion processing with thermal similarity bias on the candidate historical interval set to obtain the initial power prediction result.

[0052] S106. Substitute the initial power prediction results into the dynamic energy balance equation for energy balance back-substitution verification and correction to obtain the final photovoltaic power prediction results.

[0053] In this optional embodiment, the multi-source heterogeneous measurement data includes: high-frequency channel data, mid-frequency channel data, and low-frequency channel data; wherein, the high-frequency channel data includes at least one of global horizontal irradiance data, direct normal irradiance data, or diffuse horizontal irradiance data collected within a preset sampling period; the mid-frequency channel data includes at least one of photovoltaic module temperature data or AC power output data collected within a preset sampling period; and the low-frequency channel data includes at least one of ambient temperature data, wind speed data, or wind direction data collected within a preset sampling period.

[0054] In this optional embodiment, the continuous-time interpolation processing of multi-source heterogeneous measurement data to construct a continuous-time input path function includes: taking the current sampling time as a node, acquiring discrete observation values ​​of high-frequency channel data, mid-frequency channel data, and low-frequency channel data, and constructing corresponding discrete observation sequences; constructing a cubic polynomial as a continuous function for the discrete observation values ​​of each adjacent interval in the discrete observation sequence; constructing a tridiagonal linear equation system based on node interpolation conditions, first-order derivative continuity conditions, second-order derivative continuity conditions, and natural boundary conditions, and solving the tridiagonal linear equation system to obtain the node second-order derivative values; calculating the cubic spline coefficients in the cubic polynomial based on the node second-order derivative values ​​to determine the continuous-time interpolation function of all channel data; combining the continuous-time interpolation functions of all channel data into a meteorological input vector, and incorporating the time parameter as an additional coordinate into the meteorological input vector to obtain the continuous-time input path function.

[0055] In this optional embodiment, the dynamic energy balance equation includes a heat transfer term and an electrical power output term; wherein, the heat transfer term includes the absorbed radiative heat power of the photovoltaic module, long-wave radiation heat loss, and convective heat loss.

[0056] In this optional embodiment, the training of the physical neural differential model includes the following steps: freezing the dual-scale branch and training the neural network model set using a composite loss function; freezing the neural network model set and training the dual-scale branch using a back-substitution check consistency loss function; and training the neural network model set and the dual-scale branch respectively using a differentiated learning rate method.

[0057] In this optional embodiment, the neural network model set includes an anisotropic sky radiation correction neural network, an absorptivity correction neural network, a convective heat transfer coefficient prediction neural network, and a power conversion neural network; wherein, the anisotropic sky radiation correction neural network is used to calculate the irradiance of the tilted surface; the absorptivity correction neural network is used to calculate the effective optical absorptivity and calculate the absorbed radiative heat power based on the irradiance of the tilted surface; the convective heat transfer coefficient prediction neural network is used to calculate the convective heat loss; and the power conversion neural network is used to calculate the electrical power output term.

[0058] In this optional embodiment, calculating the continuous-time temperature trajectory and heat flux trajectory of the photovoltaic module includes: obtaining the initial temperature of the photovoltaic module at the initial moment within a given time window; using the initial temperature as the initial condition and combining it with the continuous-time input path function, numerically integrating the dynamic energy balance equation to obtain the continuous-time temperature trajectory and the time derivative corresponding to the continuous-time temperature trajectory; and calculating the heat flux trajectory by substituting the absorbed radiative heat power, long-wave radiative heat loss, convective heat loss, and electrical power output based on the continuous-time temperature trajectory and the continuous-time input path function.

[0059] In this optional embodiment, the extraction of thermal transient features from the continuous-time temperature trajectory and heat flux trajectory includes: based on the continuous-time temperature trajectory and heat flux trajectory, extracting the continuous thermal trajectory at the end of the current prediction interval to reconstruct the temperature features, as well as the maximum value of the absolute value of the photovoltaic module temperature change rate, the average convective heat loss features, the irradiance-temperature hysteresis features, and the equivalent thermal time constant features within the interval.

[0060] In this optional embodiment, the step of calculating the net thermal driving direction parameters and energy closure residuals for each prediction interval and each historical interval based on the thermal transient feature vector, and performing energy balance consistency screening for each historical interval in combination with a preset energy balance residual threshold, includes the following steps: extracting the thermal transient feature vector for each prediction interval based on the continuous time temperature trajectory and heat flux trajectory, and calculating the net thermal driving direction parameters and energy closure residuals for each prediction interval and each historical interval respectively;

[0061] The expression for the net thermal drive direction parameter is as follows:

[0062] ;

[0063] In the formula, Let be the net thermal drive direction parameter for the k-th prediction interval, and , For symbolic functions, For a moment Absorbed radiant heat power, For a moment Long-wave radiation heat loss, For a moment Convective heat loss, Let be the electrical power output at time t. Let t be the end time of the k-th prediction interval. Let k be the start time of the k-th prediction interval. Let be the duration of the k-th prediction interval;

[0064] The expression for the energy closure residual is as follows:

[0065] ;

[0066] In the formula, For the first Energy closure residuals within each prediction interval For the heat capacity of photovoltaic modules, Reconstructing temperature with respect to time for continuous thermal trajectory The trace components;

[0067] By comparing the net thermal drive direction parameters of each prediction interval and each historical interval, the historical interval that matches the net thermal drive direction parameters of the prediction interval is identified.

[0068] Based on a preset energy balance residual threshold, the energy closure residuals of the matched historical intervals are filtered, and historical intervals with energy closure residuals less than or equal to the energy balance threshold are selected.

[0069] In this optional embodiment, the step of using dual-scale branching to perform cross-attention calculation and fusion processing with thermal similarity bias on the candidate historical interval set to obtain the initial power prediction result includes:

[0070] Using a multilayer perceptron, the thermal transient feature vectors of the prediction interval and the corresponding candidate historical interval are mapped to a multidimensional latent space, and the Euclidean distance between the mapped thermal transient feature vectors of the prediction interval and the thermal transient feature vectors of the candidate historical interval is calculated to construct a physical bias matrix.

[0071] In the dual-scale branch, candidate historical interval subsets at different time scales are selected, and multi-head cross-attention scores are calculated on the key values ​​in the candidate historical interval subsets using the physical bias matrix. The attention score calculation results are then spliced ​​and linearly projected to obtain the micro-context and macro-context. The dual-scale branch includes a micro-scale branch and a macro-scale branch.

[0072] By combining the temperature change rate characteristics, the micro and macro contexts are adaptively fused to obtain the fused context. The fused context is then processed sequentially with a feedforward network, layer normalization, and linear projection to obtain the initial power prediction results.

[0073] In this optional embodiment, the step of substituting the initial power prediction result into the dynamic energy balance equation and performing energy balance back-substitution verification correction to obtain the final photovoltaic power prediction result includes: substituting the initial power prediction result into the dynamic energy balance equation to calculate the back-substitution temperature trajectory; constructing a back-substitution thermal transient feature vector based on the back-substitution temperature trajectory and calculating the back-substitution verification consistency loss function; comparing the back-substitution verification consistency loss function with a preset back-substitution verification consistency threshold to obtain the final photovoltaic power prediction result.

[0074] In this optional embodiment, the expression for the continuous-time input path function is:

[0075] ;

[0076] In the formula, For the first Each channel in continuous time The interpolation function value on, The total number of channels. For continuous-time input path functions.

[0077] In this optional embodiment, the expression for the dynamic energy balance equation is:

[0078] ;

[0079] In the formula, The heat capacity of the photovoltaic module; For a moment The temperature of the photovoltaic module; For a moment The absorbed radiant heat power; For a moment Long-wave radiation heat loss; For a moment Convectional heat loss; For a moment The electrical power output.

[0080] In this optional embodiment, the expression for the continuous-time temperature trajectory is:

[0081] ;

[0082] In the formula, For a moment The continuous-time temperature trajectory; The window start time The measured initial temperature of the photovoltaic module; The set of trainable parameters for all neural networks. ; For continuous-time input path functions, Reconstruct the trajectory for continuous temperature over time.

[0083] In this optional embodiment, the expression for the thermal transient feature vector is:

[0084] ;

[0085] In the formula, For the first The thermal transient feature vectors for each prediction interval are used to characterize the short-time thermal response state of the photovoltaic module within that prediction interval, and ; End of interval Reconstructing temperature characteristics from continuous thermal trajectories at any given time; The maximum value of the absolute value of the component temperature change rate within the interval is a characteristic feature. The average convective heat loss characteristics within the interval; For the first Irradiance-temperature hysteresis characteristics within a prediction interval; For the first The equivalent thermal time constant characteristics within each prediction interval.

[0086] Figure 2 An embodiment of an ultra-short-term photovoltaic power prediction system based on thermal state sensing according to the present invention is shown.

[0087] In this optional embodiment, the ultra-short-term photovoltaic power prediction system based on thermal state sensing includes:

[0088] The continuous path construction module 201 is used to acquire multi-source heterogeneous measurement data of photovoltaic power plants and perform continuous-time interpolation processing on the multi-source heterogeneous measurement data to construct a continuous-time input path function.

[0089] The physical information calculation module 202 accepts the continuous-time input path function from the continuous path construction module and the initial power prediction result from the thermal state perception and prediction module. It is used to input the continuous-time input path function into the pre-trained physical neural differential model to calculate the continuous-time temperature trajectory and heat flux trajectory of the photovoltaic module. The physical neural differential model includes a dynamic energy balance equation, a neural network model set, and a dual-scale branch.

[0090] The thermal transient feature extraction module 203 is connected to the physical information thermodynamics calculation module and is used to extract thermal transient features from the continuous time temperature trajectory and heat flux trajectory in each prediction interval to construct a thermal transient feature vector.

[0091] The energy balance screening module 204 accepts the thermal transient feature vector from the thermal transient feature extraction module, and uses it to calculate the net thermal driving direction parameter and energy closure residual for each prediction interval and each historical interval based on the thermal transient feature vector. It also combines the preset energy balance residual threshold to perform energy balance consistency screening for each historical interval and construct a candidate historical interval set.

[0092] The thermal state perception and prediction module 205 is connected to the energy balance screening module and is used to perform cross-attention calculation and fusion processing of thermal similarity bias on the candidate historical interval set using dual-scale branching to obtain the initial power prediction result.

[0093] The energy balance back-substitution verification and correction module 206 is connected to the physical information thermodynamics calculation module and is used to substitute the initial power prediction result into the dynamic energy balance equation to perform energy balance back-substitution verification and correction, and output the final photovoltaic power prediction result.

[0094] The continuous path construction module, physical information thermodynamics calculation module, thermal transient feature extraction module, energy balance screening module, thermal state perception prediction module, and energy balance back-substitution verification and correction module form a closed loop to achieve the prediction of photovoltaic power.

[0095] To facilitate understanding of the above technical solutions of the present invention, the following further explains the above technical solutions of the present invention from the perspective of architecture and principle, as follows:

[0096] like Figure 4 As shown, multi-source heterogeneous measurement data of photovoltaic power station are collected. High-frequency channel data is collected at the second level (the sampling period is preferably 1 second in this embodiment), and medium-frequency channel data (the sampling period is preferably 1 minute in this embodiment) and low-frequency channel data (the sampling period is preferably 5 minutes in this embodiment) are collected at the minute level.

[0097] The high-frequency channel data includes: global horizontal irradiance I, in units of Direct normal irradiance The unit is Scattered horizontal irradiance The unit is .

[0098] The intermediate frequency channel data includes: temperature data of photovoltaic modules. The unit is K, representing AC power output data. The unit is W.

[0099] Low-frequency channel data includes: ambient temperature The unit is K, wind speed The unit is and wind direction The unit is rad.

[0100] For each channel of data, obtain its discrete observation value at the sampling time node, and construct a discrete observation sequence, the expression of which is:

[0101] ;

[0102] In the formula, For the first The first channel Each sampling time, in units of ; This is the observation value corresponding to this sampling time, and the unit is the same as the first observation. The physical quantities in each channel are in the same unit; For the first The total number of sampling points for each channel within the current time window is reduced by 1.

[0103] For each adjacent interval of observations, a cubic polynomial is constructed, the expression of which is:

[0104] ;

[0105] In the formula, and The first The first channel in the Cubic spline coefficients over each sampling interval; For the first Each channel in continuous time The interpolation function value on, and the unit is the same as Consistent, This represents the n+1 sampling time of the i-th channel.

[0106] Construct a tridiagonal linear system of equations based on the point interpolation condition, the first derivative continuity condition, the second derivative continuity condition, and the natural boundary condition. The expression for the natural boundary condition is:

[0107] ;

[0108] In the formula, For the first The second derivative of the interpolation function of each channel with respect to time, and the unit is the physical quantity of the corresponding channel per square second;

[0109] In addition, , represents the interval between adjacent sampling times, in seconds (S).

[0110] set up , representing a node If the second derivative value at a given point is given, then the internal nodes satisfy a tridiagonal linear system of equations, the expression of which is:

[0111] ;

[0112] In the formula, For nodes The value of the second derivative, The sampling time interval This is the observation value corresponding to this sampling time. ,and These correspond to natural boundary conditions, This refers to the sampling time interval of the (n-1)th sampling interval in the i-th data channel. For the i-th data channel, node The second derivative of the interpolation function at that point is... For the i-th data channel, the sampling time Corresponding observations .

[0113] After solving the tridiagonal linear system of equations, the values ​​of the second derivatives at all nodes can be obtained. .

[0114] In seeking After that, the The cubic spline coefficients over each sampling interval are calculated using the following formula:

[0115] ;

[0116] In the formula, and The first The first channel in the cubic spline coefficients over each sampling interval.

[0117] After performing the above process on all channel data, all continuous-time interpolation functions are obtained, which are combined into a meteorological input vector, and the time is also... Incorporating additional coordinates, we obtain the continuous-time input path function, whose expression is:

[0118] ;

[0119] In the formula, For the first Each channel in continuous time The interpolation function value on, The total number of channels. For continuous-time input path functions.

[0120] After obtaining the continuous-time input path function, the continuous-time temperature trajectory and heat flux trajectory of the photovoltaic module are calculated using the trained physical neural differential model.

[0121] It should be noted that the full name of the physical neural differential model is the physical information neural control differential equation model, and this model consists of a physical information neural control differential equation module (composed of a dynamic energy balance equation and a neural network model) and a prediction module (composed of a dual-scale branch architecture).

[0122] Specifically, the physical neural differential model takes the dynamic energy balance equation of photovoltaic modules as its core, and its expression is:

[0123] ;

[0124] In the formula, The heat capacity of a photovoltaic module, in units of ; For a moment The temperature of the photovoltaic module, in units of ; For a moment Absorbed radiant heat power, in units of ; For a moment Long-wave radiation heat loss, in units of , For a moment Convection heat loss, in units of ; For a moment The electrical power output is measured in watts (W). This includes the absorbed radiant heat power (i.e., the absorbed radiant heat power from the sun). The calculation process is as follows:

[0125] ;

[0126] In the formula, The surface area of ​​the photovoltaic module is expressed in units of... For effective optical absorption, This represents the irradiance of the inclined surface. Among them, The calculation expression is:

[0127] ;

[0128] In the formula, For the Sigmoid function; The nominal optical absorption rate is dimensionless. The absorptivity correction range boundary is set to 0.1, constraining the correction amount within... nearby Within the range; Neural network for absorption rate correction.

[0129] It should be noted that, Output Activation function restrictions Within the range; for The input feature vector, where, This is the atmospheric mass number (dimensionless). The cumulative rainfall is expressed in mm and is used as a proxy indicator for the degree of dust accumulation. Let be the angle of incidence of sunlight on the surface of the photovoltaic module at time t, expressed in radians (rad).

[0130] The calculation expression is:

[0131] ;

[0132] In the formula, The direct normal irradiance at time t, in units of The diffuse horizontal irradiance at time t, in units of The global horizontal irradiance at time t, in units of Where is the geometric transformation factor, The angle of incidence of sunlight on the surface of the component is expressed in rad. The solar zenith angle is measured in rad. The component tilt angle, in units of Ground reflectance, dimensionless; A neural network for correcting anisotropic sky radiation.

[0133] It should be noted that, Used to compensate for errors introduced by the isotropic sky assumption; This is the input feature vector of the network;

[0134] in, The clear sky index, Radiance outside the atmosphere, in units of , Let t be the solar zenith angle at time t, representing the angle between the sun's rays and the local vertical direction, expressed in radians (rad). Let be the angle of incidence of sunlight on the surface of the photovoltaic module at time t, representing the angle between the sunlight and the normal to the module surface, in radians (rad).

[0135] Long-wave radiation heat loss The calculation process is as follows:

[0136] ;

[0137] In the formula, The Stefan-Boltzmann constant has a value of [value missing]. The effective emissivity of the component surface is dimensionless. The ambient temperature at time t is expressed as the fourth power, in Kelvin. Let t be the fourth power of the absolute temperature of the photovoltaic module.

[0138] It should be noted that this item follows known physical laws and does not require neural network correction.

[0139] Convection heat loss The calculation process is as follows:

[0140] ;

[0141] In the formula, The output of the neural network for predicting convective heat transfer coefficient is ensured to be non-negative by applying a softplus activation function; Let be the input feature vector of this network. Wind speed (unit: ), Wind direction (in rad) relative to the component's normal direction. Multiplied by temperature difference ( This ensures that convective heat loss is zero under thermal equilibrium conditions.

[0142] Electric power output The calculation process is as follows:

[0143] ;

[0144] In the formula, It is a power conversion neural network, and the output is activated by the ReLU function to ensure non-negativity. It is used to describe nonlinear photoelectric conversion characteristics, including low irradiance efficiency degradation and maximum power point tracking behavior.

[0145] It should be noted that, and For the feedforward fully connected networks constructed to correspond to the physical subtasks, a multilayer perceptron structure is preferred. It employs three hidden layers, each with 64 neurons, and linear activation of the output layer; It uses two hidden layers, each with 32 neurons, and an output layer. activation; It employs three hidden layers, each with 64 neurons, and softplus activation of the output layer; The network employs three hidden layers, each with 128 neurons, and the output layer is activated using ReLU. Each hidden layer uses the ReLU activation function. Its specific architecture is as follows:

[0146]

[0147]

[0148]

[0149]

[0150] In the formula, the symbol For the two adjacent floors by Dimensional input mapping to Dimensional output, for example: .

[0151] also, The output layer uses a linear activation function to allow the anisotropic radiation correction to be positive or negative; The output layer uses the tanh activation function to limit the absorption rate correction within a bounded range; The output layer uses the softplus activation function to ensure that the convective heat transfer coefficient is non-negative; The output layer uses the ReLU activation function to ensure that the electrical power output is non-negative.

[0152] The initial temperature at the initial moment within a given time window. Using the initial conditions, the Dormand-Prince adaptive step-size method is used to numerically integrate and solve the dynamic energy balance equation, obtaining the continuous-time temperature trajectory, whose expression is:

[0153] ;

[0154] In the formula, For a moment The continuous-time temperature trajectory, in units of ; The window start time The measured initial temperature of the photovoltaic module, in units of ; The set of trainable parameters for all neural networks. ; For continuous-time input path functions, Reconstruct the trajectory for continuous temperature over time.

[0155] For each prediction interval , and The first The start and end times of each prediction interval. For the first The duration of each prediction interval, in seconds.

[0156] Based on the continuous-time temperature trajectory and heat flux trajectory, thermal transient features are extracted to construct a thermal transient feature vector. Its expression is:

[0157] ;

[0158] In the formula, For the first The thermal transient feature vectors for each prediction interval are used to characterize the short-time thermal response state of the photovoltaic module within that prediction interval, and ; End of interval Reconstructing temperature characteristics from continuous thermal trajectories at time points, in units of ; This represents the maximum absolute value of the component's temperature change rate within the interval, in units of... The average convective heat loss characteristic within the interval, in units of ; For the first The irradiance-temperature hysteresis characteristic within each prediction interval, in units of ; For the first The equivalent thermal time constant characteristics within each prediction interval, in units of .

[0159] in, The definition expression is:

[0160] ;

[0161] In the formula: For interval The domain is a continuous-time variable. ; For a moment The first derivative of the temperature at which the continuous trajectory is reconstructed, in units of .

[0162] The definition expression is:

[0163] ;

[0164] In the formula: For a moment The convective heat loss is calculated from the temperature reconstructed from the continuous thermal trajectory, and the unit is W.

[0165] The definition expression is:

[0166] ;

[0167] In the formula, For a moment Irradiance of the inclined surface at the location, in units of The irradiance of the inclined surface with respect to time The trace components; Reconstructing temperature with respect to time for continuous thermal trajectory The trace components; For the first The irradiance-temperature hysteresis characterization quantity within a prediction interval is used to characterize the degree of hysteresis between changes in tilted surface irradiance and component temperature response. The definition expression is: ;

[0168] In the formula, For the first The average effective convective heat transfer coefficient within each prediction interval, in units of Its definition expression is:

[0169] ;

[0170] In the formula, The heat capacity of the photovoltaic module is expressed in units of 1. For the first The equivalent thermal time constant within each prediction interval is used to characterize the thermal response speed of the component. A neural network for predicting convective heat transfer coefficients.

[0171] Calculate the net thermal drive direction parameters for the prediction interval and the historical interval respectively. and energy closure residual Only those meeting the requirements of consistent net heat drive direction and energy closure residual not exceeding the threshold will be retained. The historical intervals are used to form a set of candidate historical intervals for each future prediction interval. .

[0172] Specifically, first calculate the first... Net thermal drive direction parameters for each prediction interval Its expression is:

[0173] ;

[0174] In the formula, This is a dimensionless net heat drive direction indicator. For sign functions; when When the net heat drive of the interval is positive, When the net heat drive of the interval is negative, At this point, the interval is approximately in equilibrium due to net heat drive. For a moment The absorbed radiant heat power represents the heat input power corresponding to the solar radiation absorbed by the photovoltaic module, and its unit is W. For a moment Long-wave radiation heat loss represents the heat power lost by the photovoltaic module to the outside world through thermal radiation, measured in W. For a moment Convective heat loss represents the heat power dissipated by the photovoltaic module through convection heat exchange with the surrounding air, measured in W. For a moment The electrical power output represents the electrical power output of the photovoltaic module at that moment, measured in W.

[0175] Calculate the next Energy closure residuals for each prediction interval Its expression is:

[0176] ;

[0177] In the formula, For the first The energy balance residuals within each prediction interval, expressed in W, characterize the degree of consistency between the thermal trajectory and the dynamic energy balance equation. Reconstructing temperature with respect to time for continuous thermal trajectory The trace amount, This refers to the heat capacity of a photovoltaic module.

[0178] For the future The set of candidate historical intervals for each prediction interval is defined as follows:

[0179] ;

[0180] In the formula, For the future The set of candidate historical intervals corresponding to each prediction interval; and They are respectively the historical number Net heat drive direction index and energy balance residual for each interval; For the future Net heat drive direction indicator for each prediction interval; This is a preset energy balance residual threshold, in W, used to filter out historical intervals with inconsistent energy balance.

[0181] like Figure 5 As shown, the historical input sequence and future input sequence Each network is mapped to an embedding sequence through variable selection. and The historical embeddings are encoded by a temporal convolutional network and bidirectional gated recurrent units to obtain the historical encoded sequence. .

[0182] Constructing a physical bias matrix using a multilayer perceptron Its expression is:

[0183] ;

[0184] In the formula, For the future Thermal transient feature vectors for each prediction interval; For the first time in history Thermal transient feature vectors of each interval; The square of the Euclidean distance; To map the thermal transient feature vector to The architecture of a two-layer multilayer perceptron with a 3D latent space is as follows: It should be noted that the hidden layer uses the ReLU activation function, while the output layer uses the linear activation function.

[0185] The multi-head cross-attention score is calculated by subtracting from the physical bias matrix, specifically as follows:

[0186] for The first attention head For each key-value pair in the candidate historical interval set, cross-attention is calculated using the following expression:

[0187] ;

[0188] In the formula, For the first Attention score matrix for each attention head; Query matrix, by future embedded sequence Obtained by linear projection; The key matrix consists of the encoded sequences of candidate historical intervals. Obtained by linear projection; Number of candidate historical intervals; The key dimension for each attention head; This is a learnable scaling factor. Subtract... This results in candidate historical intervals with significant differences in thermal transient features receiving lower attention weights.

[0189] The output of each attention head is ,in For a matrix, For the value dimension, all attention heads are concatenated and linearly projected to obtain the cross-attention output. Cross-attention with thermal similarity bias is calculated for each historical interval using a dual-scale branch, which includes a micro-scale branch and a macro-scale branch. The micro-scale branch focuses on the most recent... Perform cross-attention calculations on the thermal similarity bias for each candidate historical interval to obtain the micro-context. Used to capture the rapid thermal transients caused by recent cloud transits; macroscale branch for past Perform cross-attention calculation with the thermal similarity bias on each candidate historical interval to obtain the macro-context. It is used to provide long-term background information on the evolution of weather systems; For the embedded dimension. Through fusion gating The adaptive fusion of macroscopic and microscopic contexts is expressed as follows:

[0190] ;

[0191] In the formula, This is a concatenation operation along the feature dimension; The temperature change rate embedding matrix is ​​the future temperature change rate predicted by the CDE (i.e., the physical information neural control differential equation module) at each time step. Obtained through linear mapping; The weight matrix is ​​a learnable weight matrix; For learnable bias vectors, For length is A vector of all 1s; This is the Sigmoid function. When the rate of temperature change is large, When the temperature approaches 1, the fusion result tends towards micro-branching; when the temperature change rate is small, Approaching 0, the fusion result leans towards the macroscopic branch.

[0192] Furthermore, the calculation method for the fusion context is as follows:

[0193] ;

[0194] In the formula, This is element-wise multiplication (Hadamard product). For micro context, For the macro context, For context fusion.

[0195] After the fusion context is processed by a feedforward network and layer normalization, the initial power prediction result is obtained through linear projection, and its expression is:

[0196] ;

[0197] In the formula, This is the initial power prediction result; This is the output of the feedforward network; To output the projected weight vector; This is the output bias vector; For the first The initial power prediction value for each prediction time step, in W.

[0198] Substituting the initial power prediction results into the dynamic energy balance equation and solving it yields the back-substitution temperature trajectory. Its expression is:

[0199] ;

[0200] In the formula, The component temperature trajectory obtained from the back-substitution verification, in units of , and These represent the absorbed radiative heat power, long-wave radiative heat loss, and convective heat loss calculated based on the back-substitution temperature trajectory, respectively, with units of [missing information]. For the discrete initial power prediction sequence The constructed piecewise constant continuous power function is expressed in W.

[0201] Construct a transient eigenvector of the back-substitution temperature trajectory. And calculate the back-substitution verification consistency loss function, the expression of which is:

[0202] ;

[0203] In the formula: The first [item] is calculated based on the replacement temperature trajectory and the replacement heat flux trajectory. Each iteration of the thermal transient eigenvectors; For the future The target thermal transient feature vector for each prediction interval; This is the loss function for back-substitution verification consistency.

[0204] Based on the consistency loss function of back-substitution verification and the preset threshold The comparison results are used to evaluate the initial power prediction results, and the final power prediction result is obtained, which is expressed as follows:

[0205] ;

[0206] In the formula, This is the final power prediction result; This is the initial power prediction result; The preset consistency threshold for back-substitution verification; To check and correct the step size; To back-substitute and verify the gradient of the consistency loss with respect to the initial power prediction results. When When the initial power prediction result and the thermal transient characteristics of the back-substitution have met the preset consistency requirements, no back-substitution verification correction is performed, and the result is directly applied. As the final power prediction result, it is output. The initial power prediction results are then corrected as follows:

[0207] ;

[0208] In the formula, The initial power prediction results are as follows. The gradient of the consistency loss with respect to the initial power prediction results is used for back-substitution verification.

[0209] It should be noted that the training of the physical neural differential model adopts a three-stage training strategy, specifically:

[0210] like Figure 6 As shown, in the first stage, a composite loss function is used to train the neural network model set, and its expression is:

[0211] ;

[0212] In the formula, This is the composite loss function of the physical information neural control differential equation module. This is the root mean square error loss for temperature. This is the power root mean square error loss. This is the energy balance regularization term. and These are the dimensionless weighting coefficients for power loss and energy balance loss, respectively. It should be noted that... More preferably More preferably, it should be 0.1. In the first stage, the AdamW optimizer is preferred, with a learning rate set to... .

[0213] The root mean square error loss for temperature is defined as follows:

[0214] ;

[0215] The power root mean square error loss is defined as follows:

[0216] ;

[0217] The energy balance regularization term is defined as follows:

[0218] ;

[0219] In the formula, For monitoring time The measured component temperature at the location, in units of For monitoring time The measured power output at the location is expressed in units of... and These represent the number of sample points involved in calculating temperature loss, power loss, and energy balance loss within a training batch. In the second stage, the neural network model set is frozen, and only the parameters of the dual-scale branch (i.e., the thermal state sensing prediction module) are trained. And introduce back-substitution verification consistency loss. Furthermore, the AdamW optimizer is preferred, with a learning rate set to [value missing]. In the third stage, all parameters of the neural network model set and the dual-scale branch are jointly adjusted using a differentiated learning rate, and the joint loss function is defined as:

[0220] ;

[0221] In the formula, The physical consistency loss weighting coefficient is preferably within the range of [value range missing]. More preferably The preferred value range for the consistency loss weighting coefficient during back-substitution verification is [value range missing]. Differential learning rate refers to the rate applied to the set of parameters of a neural network model. and dual-scale branch parameters Different learning rates were used, among which:

[0222] ;

[0223] In the formula, The learning rate for the set of neural network models in the third stage; The learning rate for the dual-scale branch in the third stage; both satisfy... This helps to suppress excessive drift in the parameters of the neural network model set during the joint fine-tuning stage. Furthermore, a cosine annealing learning rate scheduling strategy is employed during training, with the minimum learning rate set to... .

[0224] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0225] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0226] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0227] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0228] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0229] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A method for predicting ultra-short-term photovoltaic power based on thermal state sensing, characterized in that, include: Acquire multi-source heterogeneous measurement data from photovoltaic power plants, and perform continuous-time interpolation on the multi-source heterogeneous measurement data to construct a continuous-time input path function; The continuous-time input path function is input into a pre-trained physical neural differential model to calculate the continuous-time temperature trajectory and heat flux trajectory of the photovoltaic module; wherein, the physical neural differential model includes a dynamic energy balance equation, a neural network model set, and a two-scale branch; Within each prediction interval, thermal transient features are extracted from the continuous-time temperature trajectory and heat flux trajectory to construct a thermal transient feature vector; Based on the thermal transient feature vector, the net thermal driving direction parameters and energy closure residuals for each prediction interval and each historical interval are calculated respectively. Combined with the preset energy balance residual threshold, the energy balance consistency screening process is carried out for each historical interval to construct a candidate historical interval set. By using a dual-scale branch, similarity bias cross-attention calculation and fusion processing are performed on the candidate historical interval set to obtain the initial power prediction results; Substitute the initial power prediction results into the dynamic energy balance equation, perform energy balance back-substitution verification and correction, and obtain the final photovoltaic power prediction results.

2. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 1, characterized in that, The multi-source heterogeneous measurement data includes: high-frequency channel data, mid-frequency channel data, and low-frequency channel data; The high-frequency channel data includes at least one of global horizontal irradiance data, direct normal irradiance data, or scattered horizontal irradiance data collected within a preset sampling period. The intermediate frequency channel data includes at least one of the photovoltaic module temperature data or AC power output data collected within a preset sampling period; The low-frequency channel data includes at least one of the ambient temperature data, wind speed data, or wind direction data collected within a preset sampling period.

3. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 1, characterized in that, The step of performing continuous-time interpolation on multi-source heterogeneous measurement data to construct a continuous-time input path function includes the following steps: Using the current sampling time as a node, obtain discrete observation values ​​of high-frequency channel data, mid-frequency channel data, and low-frequency channel data, and construct the corresponding discrete observation sequence; For each adjacent interval of the discrete observation value in the discrete observation sequence, construct a cubic polynomial as a continuous function; A tridiagonal linear equation system is constructed based on nodal interpolation conditions, first derivative continuity conditions, second derivative continuity conditions, and natural boundary conditions. The tridiagonal linear equation system is then solved to obtain the nodal second derivative values. Based on the second derivative values ​​of the nodes, the cubic spline coefficients in the cubic polynomial are calculated to determine the continuous-time interpolation function for all channel data. The continuous-time interpolation functions of all channel data are combined into a meteorological input vector, and the time parameter is used as an additional coordinate to be incorporated into the meteorological input vector to obtain the continuous-time input path function.

4. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 1, characterized in that, The dynamic energy balance equation includes: a heat transfer term and an electrical power output term; The heat exchange items include the absorbed radiative heat power of the photovoltaic module, long-wave radiation heat loss, and convective heat loss.

5. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 1, characterized in that, The training of the physical neural differential model includes the following steps: Freeze the dual-scale branches and train the neural network model set using a composite loss function; The neural network model set was frozen, and the dual-scale branch was trained using the back-substitution verification consistency loss function; The differential learning rate method was used to train the neural network model set and the dual-scale branch respectively.

6. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 1, characterized in that, The neural network model set includes: an anisotropic sky radiation correction neural network, an absorptivity correction neural network, a convective heat transfer coefficient prediction neural network, and a power conversion neural network. The anisotropic sky radiation correction neural network is used to calculate the irradiance of the tilted surface. The absorptivity correction neural network is used to calculate the effective optical absorptivity and to calculate the absorbed radiant heat power based on the irradiance of the inclined surface. The convective heat transfer coefficient prediction neural network is used to calculate convective heat loss; The power conversion neural network is used to calculate the electrical power output term.

7. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 1, characterized in that, The calculation of the continuous-time temperature trajectory and heat flux trajectory of the photovoltaic module includes the following steps: Based on a given time window, obtain the initial temperature of the photovoltaic module at the initial moment within the time window; Using the initial temperature as the initial condition and combining the continuous-time input path function, the dynamic energy balance equation is solved by numerical integration to obtain the continuous-time temperature trajectory and the corresponding time derivative of the continuous-time temperature trajectory. Based on the continuous-time temperature trajectory and the continuous-time input path function, the heat flux trajectory is calculated by substituting the absorbed radiative heat power, long-wave radiative heat loss, convective heat loss, and electrical power output.

8. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 7, characterized in that, The extraction of thermal transient features from continuous-time temperature and heat flux trajectories includes: Based on the continuous time temperature trajectory and heat flux trajectory, the continuous thermal trajectory at the end of the current prediction interval is extracted to reconstruct the temperature characteristics, as well as the maximum value of the absolute value of the temperature change rate of the photovoltaic module, the average convective heat loss characteristics, the irradiation-temperature hysteresis characteristics, and the equivalent thermal time constant characteristics within the interval.

9. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 1, characterized in that, The process of calculating the net thermal driving direction parameters and energy closure residuals for each prediction interval and each historical interval based on thermal transient feature vectors, and then performing energy balance consistency screening for each historical interval in conjunction with a preset energy balance residual threshold, includes the following steps: Based on the thermal transient feature vector, the net thermal driving direction parameters and energy closure residuals for each prediction interval, as well as the net thermal driving direction parameters and energy closure residuals for each historical interval, are calculated respectively. The expression for the net thermal drive direction parameter is as follows: ； In the formula, Let be the net thermal drive direction parameter for the k-th prediction interval, and , For symbolic functions, For a moment Absorbed radiant heat power, For a moment Long-wave radiation heat loss, For a moment Convective heat loss, Let be the electrical power output at time t. Let t be the end time of the k-th prediction interval. Let k be the start time of the k-th prediction interval. Let be the duration of the k-th prediction interval; The expression for the energy closure residual is as follows: ； In the formula, For the first Energy closure residuals within each prediction interval For the heat capacity of photovoltaic modules, Reconstructing temperature with respect to time for continuous thermal trajectory The trace components; By comparing the net thermal drive direction parameters of each prediction interval and each historical interval, the historical interval that matches the net thermal drive direction parameters of the prediction interval is identified. Based on a preset energy balance residual threshold, the energy closure residuals of the matched historical intervals are filtered, and historical intervals with energy closure residuals less than or equal to the energy balance threshold are selected.

10. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 1, characterized in that, The method of using dual-scale branching to perform similarity-biased cross-attention calculation and fusion processing on the candidate historical interval set to obtain the initial power prediction result includes the following steps: Using a multilayer perceptron, the thermal transient feature vectors of the prediction interval and the corresponding candidate historical interval are mapped to a multidimensional latent space, and the Euclidean distance between the thermal transient feature vectors of the prediction interval and the candidate historical interval after mapping is calculated to construct a physical bias matrix. In the dual-scale branch, candidate historical interval subsets at different time scales are selected, and multi-head cross-attention scores are calculated on the key values ​​in the candidate historical interval subsets using the physical bias matrix. The attention score calculation results are then spliced ​​and linearly projected to obtain the micro-context and macro-context. The dual-scale branch includes a micro-scale branch and a macro-scale branch. By combining the temperature change rate characteristics, the micro-context and macro-context are adaptively fused to obtain the fused context; The fusion context is processed sequentially using a feedforward network, layer normalization, and linear projection to obtain the initial power prediction results.

11. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 1, characterized in that, The process of substituting the initial power prediction result into the dynamic energy balance equation and performing energy balance back-substitution verification and correction to obtain the final photovoltaic power prediction result includes the following steps: Substituting the initial power prediction results into the dynamic energy balance equation, the back-substitution temperature trajectory is calculated. Construct a back-substitution thermal transient feature vector based on the back-substitution temperature trajectory, and calculate the back-substitution verification consistency loss function; The back-substitution verification consistency loss function is compared with the preset back-substitution verification consistency threshold to obtain the final photovoltaic power prediction result.

12. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 3, characterized in that, The expression for the continuous-time input path function is: ； In the formula, For the first Each channel in continuous time The interpolation function value on, The total number of channels. For continuous-time input path functions.

13. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 4, characterized in that, The expression for the dynamic energy balance equation is: ； In the formula, The heat capacity of the photovoltaic module; For a moment The temperature of the photovoltaic module; For a moment The absorbed radiant heat power; For a moment Long-wave radiation heat loss; For a moment Convectional heat loss; For a moment The electrical power output.

14. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 7, characterized in that, The expression for the continuous-time temperature trajectory is: ； In the formula, For a moment The continuous-time temperature trajectory; The window start time The measured initial temperature of the photovoltaic module; The set of trainable parameters for all neural networks. ; For continuous-time input path functions, Reconstruct the trajectory for continuous temperature over time.

15. The ultra-short-term photovoltaic power prediction method based on thermal state sensing according to claim 8, characterized in that, The expression for the thermal transient feature vector is: ； In the formula, For the first The thermal transient feature vectors for each prediction interval are used to characterize the short-time thermal response state of the photovoltaic module within that prediction interval, and ; End of interval Reconstructing temperature characteristics from continuous thermal trajectories at any given time; The maximum value of the absolute value of the component temperature change rate within the interval is a characteristic feature. The average convective heat loss characteristics within the interval; For the first Irradiance-temperature hysteresis characteristics within a prediction interval; For the first The equivalent thermal time constant characteristics within each prediction interval.

16. A super-short-term photovoltaic power prediction system based on thermal state sensing, characterized in that, include: The continuous path construction module is used to acquire multi-source heterogeneous measurement data from photovoltaic power plants and perform continuous-time interpolation processing on the multi-source heterogeneous measurement data to construct a continuous-time input path function. The physical information thermodynamics calculation module is used to input the continuous-time input path function into the pre-trained physical neural differential model to calculate the continuous-time temperature trajectory and heat flux trajectory of the photovoltaic module; wherein, the physical neural differential model includes a dynamic energy balance equation, a neural network model set, and a dual-scale branch; The thermal transient feature extraction module is used to extract thermal transient features from the continuous-time temperature trajectory and heat flux trajectory within each prediction interval, and construct a thermal transient feature vector. The energy balance screening module is used to calculate the net thermal driving direction parameters and energy closure residuals for each prediction interval and each historical interval based on the thermal transient feature vector, and to perform energy balance consistency screening for each historical interval in combination with the preset energy balance residual threshold, thereby constructing a set of candidate historical intervals. The thermal state perception and prediction module is used to perform similarity bias cross-attention calculation and fusion processing on the candidate historical interval set using dual-scale branching to obtain the initial power prediction result; The energy balance back-substitution verification and correction module is used to substitute the initial power prediction results into the dynamic energy balance equation, perform energy balance back-substitution verification and correction processing, and obtain the final photovoltaic power prediction results.

Images