A model predictive control method and system for smoothing photovoltaic power fluctuations

Abstract

The application discloses a model prediction control method and system for smoothing photovoltaic power fluctuation, and relates to the technical field of new energy power generation control. The method comprises the following steps: collecting historical data and meteorological data of a photovoltaic power station, and constructing a multivariate time series data set after cleaning and normalization processing; processing a non-stationary photovoltaic power sequence by using a variational mode decomposition and a reversible instance standard (RevIN); constructing a prediction model based on a TCN-iTransformer, extracting local fluctuation characteristics of mode components by using a TCN module, and capturing global dependencies among multivariate by using an inverted attention mechanism; introducing a solar orbit as a physical prior, combining a cross-attention and a physical gating strategy, and ensuring that a prediction result conforms to a physical law. Finally, a hybrid energy storage state space model is established, a high-precision prediction sequence is introduced into an MPC rolling optimization framework, optimal control instructions are solved under the condition of meeting a grid-connected fluctuation rate constraint, and effective smoothing of photovoltaic fluctuation is realized.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation and energy storage control technology, specifically relating to a hybrid energy storage predictive control method and system for photovoltaic power fluctuations, which is used to improve the stability of the output power of the photovoltaic power generation system and the coordinated operation capability of the energy storage system. Background Technology

[0002] Photovoltaic power output is significantly affected by meteorological factors such as solar irradiance, temperature, and cloud cover, exhibiting considerable randomness and volatility. When a high proportion of photovoltaic power is connected to the grid, severe power fluctuations can threaten grid frequency stability. Configuring a hybrid energy storage system (HESS), which includes energy storage units (such as lithium batteries) and power storage units (such as flywheels), and utilizing the complementary characteristics of both, is an effective means of mitigating photovoltaic power fluctuations.

[0003] The key to achieving efficient collaboration in hybrid energy storage systems lies in high-precision ultra-short-term power prediction and real-time optimized control strategies. In terms of power prediction, existing technologies mostly employ Long Short-Term Memory (LSTM) networks or conventional Transformer models. However, existing methods suffer from the following major technical problems in practical applications: First, the non-stationary distribution of data leads to prediction bias. Photovoltaic power and meteorological data are inherently non-stationary time series, affected by seasonal changes and sudden weather changes. The statistical distribution of training data and real-time data is often inconsistent (for example, summer models are difficult to predict winter data), resulting in poor generalization ability of traditional deep learning models based on the independent and identically distributed assumption, and a significant decrease in prediction accuracy over time.

[0004] Second, the multivariate features are not fully utilized. Conventional Transformer models typically embed all variables (such as irradiance, temperature, and humidity) at the same time into a single feature vector (Token). This approach focuses on capturing the time-dependent relationship but ignores the complex dynamic relationships between different physical variables (for example, the effect of temperature changes on power has a thermal inertia lag, while irradiance changes are instantaneous), resulting in insufficient robustness of the prediction model under complex weather conditions.

[0005] Third, there is a lack of physical constraint mechanisms. Purely data-driven models are prone to outputting power predictions that violate physical laws (such as "generating electricity at night") under nighttime or extreme abnormal inputs. When these unreliable prediction data are used as feedforward inputs to the model predictive control (MPC) system, they can cause the energy storage system to generate incorrect charge and discharge commands, resulting in unnecessary battery cycle losses.

[0006] Therefore, there is an urgent need for a predictive control method that can overcome data nonstationarity, effectively decouple multivariate characteristics, and integrate physical prior knowledge to improve the accuracy and economy of photovoltaic smooth control. Summary of the Invention

[0007] The present invention aims to solve the technical problems in the prior art, such as the low accuracy of photovoltaic power prediction due to the non-stationarity of data distribution, insufficient modeling of complex multivariate dependencies, lack of physical consistency of pure data-driven models, and the tendency of traditional control strategies to cause frequent operation of energy storage systems, thereby shortening their lifespan.

[0008] The first aspect of this invention provides a model-based predictive control method for smooth photovoltaic power fluctuations guided by physical priors. The overall process includes a data construction and preprocessing step, a physically guided hybrid prediction step, and a rolling time-domain smoothing control (MPC) step. These steps form a closed-loop mechanism of "data perception—physical prediction—optimized control" through data flow and control commands. The specific implementation is as follows: Step 1: Data Construction and Preprocessing.

[0009] Historical power sequences and meteorological data of photovoltaic power plants are acquired, and basic preprocessing such as anomaly cleaning and standardization is performed. At the same time, based on geographical information such as latitude and longitude of the prediction period, physical parameters of the solar trajectory such as solar altitude angle and azimuth angle are calculated to construct a multivariate time series input set containing physical prior features.

[0010] Step 2, Physically Guided Hybrid Prediction.

[0011] A multi-level feature extraction architecture of "decomposition-local extraction-global correlation" is employed to process the input set and perform ultra-short-term power prediction. Specifically, this includes: First, a multi-scale signal decomposition algorithm is used to decompose the non-stationary photovoltaic power sequence into eigenmode components of different frequencies. For example, variational mode decomposition (VMD) is used to decompose the photovoltaic power sequence into K eigenmode components with finite bandwidth and center frequency. Furthermore, the invertible instance normalization (RevIN) strategy is combined to align the distribution of each component before feature extraction and perform an inverse transformation to restore the physical dimensions before output.

[0012] Secondly, a local feature extraction module is used to capture the local short-term dependency features of each modality and meteorological variable. For example, the local feature extraction module uses a temporal convolutional network (TCN) and uses an expanded causal convolutional structure to extract the local temporal features of each normalized component as the local short-term dependency features.

[0013] Finally, a prediction model is constructed, and the processed features are input to establish a global relationship between meteorological factors and power modes along the variable dimension. That is, each power mode component and meteorological variable is mapped to an independent variable feature vector, establishing cross-variable dependencies along the variable dimension.

[0014] During the prediction process, a physical prior guidance mechanism is used to dynamically adjust the model's attention weight to historical features using the physical parameters of the solar trajectory. Furthermore, a gating strategy based on physical thresholds is employed to correct the model's initial output, generating a photovoltaic power prediction sequence. Specifically, a physical cross-attention mechanism is introduced, mapping the calculated physical parameters of the solar trajectory to a query vector, dynamically guiding the model to focus on effective meteorological features at different times, and outputting a high-precision photovoltaic power prediction sequence.

[0015] For example, the prediction model of this invention is a hybrid prediction model employing an inverted Transformer model (iTransformer). Its construction and processing include: Variable embedding: Mapping each meteorological variable and power mode component into semantically independent high-dimensional variable feature vectors; Physical prior mapping: Calculate the solar altitude angle and azimuth angle based on the latitude, longitude and timestamp of the photovoltaic power station, and map them into a physical prior feature vector aligned with the variable dimension; Physics-guided cross-attention calculation: The physical prior feature vector is used as the query vector, and the historical variable features are used as the key vector and value vector. The model’s attention weight for different meteorological variables is dynamically adjusted through cross-attention operation.

[0016] Step 3, MPC rolling time-domain smoothing control.

[0017] A state-space model incorporating the dynamic characteristics of the hybrid energy storage system is established. The photovoltaic power prediction sequence is introduced as a feedforward disturbance into the rolling optimization framework of model predictive control (MPC). An optimization objective function is constructed, constraints are set, and the optimal charging and discharging power command at the current moment is solved. The command is then sent to the converter of the hybrid energy storage system via a communication interface to drive the hybrid energy storage system's operation and smooth grid-connected power.

[0018] Unlike traditional weighted multi-objective optimization, this invention constructs a cost function with the objective of minimizing the sum of squares of the deviations between grid-connected power and the predicted photovoltaic power, and sets the power volatility at the grid connection point as a hard constraint. Within each control cycle, this constrained optimization problem is solved, ensuring that the grid-connected power follows the inherent output trajectory of the photovoltaic system as closely as possible while keeping the volatility within limits. This generates optimal power dispatch commands to control the hybrid energy storage system's operation.

[0019] Furthermore, the physical-guided hybrid prediction in step 2 of this invention also includes: correcting the initial output of the model using a gating strategy based on a physical threshold; The physical threshold-based gating strategy corrects the initial output of the model by detecting the solar elevation angle at the prediction time after the model outputs the initial prediction sequence. If the solar elevation angle is lower than the preset radiation threshold, the gating logic is forcibly triggered to correct the photovoltaic power prediction value at that time to zero, and the corrected sequence is transmitted to step 3.

[0020] According to the above technical solution, the present invention forms a closed-loop mechanism of "data perception - physical prediction - optimized control" between each step through data flow and control commands.

[0021] Furthermore, the data preprocessing step refers to the anomalies and missing data in the dataset caused by the sensor itself and the natural environment. To obtain a high-quality dataset, this embodiment performs the following hierarchical processing: Level 1: Physical rule correction. Based on prior physical knowledge, the system reads the current timestamp and latitude / longitude to calculate the real-time solar altitude angle. If it is nighttime but the sensor reading shows that the photovoltaic power is not zero, it is determined to be sensor drift. The photovoltaic power value is forcibly corrected to zero without interpolation to conform to physical laws.

[0022] Level 2: Statistical anomaly handling and KNN interpolation. For data from non-nighttime periods, statistical methods are employed. Outliers are identified in principle. For identified outliers and original missing values, the K-Nearest Neighbors (KNN) algorithm, based on neighborhood information, is used for imputation. The Euclidean distance between the sample to be imputed and samples in the historical database is calculated using the following formula: After identifying N neighbors, the average value of the neighboring samples is calculated to obtain the target value of the sample to be filled. The target value is calculated as follows: This reduces error accumulation and provides a stable initial data distribution for the subsequent RevIN module.

[0023] This repair strategy effectively reduces error accumulation caused by equipment malfunctions or short-term sampling distortions while maintaining the overall trend of the data. Secondly, to minimize the impact of dimensional differences between different features on model convergence speed and prediction accuracy, basic standardization is performed on the repaired data, improving the model's training stability and generalization ability, and providing initial data distribution consistency for subsequent Reversible Normalization (RevIN). Finally, the dataset is split, generating multiple training-test subsets through a sliding window. This partitioning method maintains the continuity of time-series data, avoids data leakage issues, and ensures the generalization reliability of the prediction model in real-world scenarios.

[0024] To address the non-stationarity and multimodal fluctuation characteristics of photovoltaic power signals, this invention constructs an ultra-short-term high-precision prediction model based on TCN-iTransformer. Considering the multi-frequency superposition and sudden fluctuation characteristics of photovoltaic output due to the coupling influence of multiple factors (irradiance, temperature, etc.), the processing flow in this embodiment strictly follows the following sequence: Step 1: Introduce the Variational Mode Decomposition (VMD) algorithm to decompose the processed historical data into several intrinsic mode components (IMFs) with different frequency characteristics. Each IMF has an independent center frequency, achieving physical isolation from high-frequency noise to low-frequency trends, thereby reducing signal complexity and laying the foundation for subsequent deep time series modeling. The objective function used for decomposition is as follows: Step Two: Considering that the distribution of photovoltaic data may significantly shift under different collection periods and meteorological conditions, to avoid a decrease in model generalization ability due to distribution differences between the training and test sets, RevIN is used with a symmetrical structure of "normalization-modeling-denormalization". In forward normalization, for each input sequence... The Each feature is used to calculate the mean and standard deviation over the time dimension at the instance level, then the data is standardized, and finally a learnable affine transformation is introduced: In inverse normalization, the output data of the intermediate model Perform inverse transform, then inverse normalization: Step 3: The standardized data of each modal component after RevIN processing is input into the TCN. The TCN extracts local temporal features using a one-dimensional dilated convolutional structure. This structure has parallel computation and a long receptive field, which can effectively capture local perturbation patterns in the power sequence and reduce short-term prediction delay. In this embodiment, the TCN adopts a grouped convolution strategy. For the K IMF components obtained from VMD decomposition, independent feature extraction channels are established to ensure that high-frequency noise features and low-frequency trend features do not interfere with each other before entering the global model.

[0025] In actual online prediction, to overcome the endpoint effect of VMD decomposition at the end of the sequence and to ensure real-time performance, this invention adopts an overlapping sliding window strategy. During each prediction, a window containing the current time and a past period (e.g., the past 24 hours) of historical data is extracted for overall decomposition. Only the IMF (Integrated Feature Function) of the window center or stable region is extracted as the feature input for the current time, thus ensuring the stability of the input TCN data.

[0026] Step 4: Next, iTransformer time series modeling is introduced for prediction. iTransformer treats the complete time series of each physical variable as an independent high-dimensional variable feature vector. It uses a self-attention mechanism to interact across variable dimensions, capturing the complex nonlinear relationships between "weather-power" and "mode-mode".

[0027] To strengthen the constraints of physical laws, this embodiment introduces a solar trajectory guidance mechanism: the solar altitude angle and azimuth angle within the prediction period are mapped into physical feature vectors, and their temporal patterns are extracted using a Transformer encoder; subsequently, feature fusion is performed through a cross-attention mechanism: the encoded prior information of the future solar position is used as the query vector, and the historical features extracted by the TCN are used as the key and value. This allows the model to dynamically adjust the attention weight of historical meteorological features based on the deterministic information of the future solar position.

[0028] This mechanism enables the model to dynamically adjust the weights of historical meteorological features (Key / Value) based on deterministic information (Query) about the future sun's position. The cross-attention structure is as follows: Step 5: Physically Gated Output. After the model generates the initial prediction sequence, it is mapped to the prediction length through a projection layer. Finally, a physical gating strategy based on the solar altitude angle is applied: the solar altitude angle at the prediction time is detected, and if it is lower than a preset radiation threshold, the predicted power value for that time is forcibly set to zero. This hard constraint completely eliminates spurious power noise at night, achieving an organic unity between historical data-driven approaches and physical mechanism constraints.

[0029] In summary, by organically integrating VMD decomposition, RevIN normalization, TCN feature extraction, and iTransformer time series modeling, this invention achieves an integrated design from multimodal feature extraction to physical constraint prediction, significantly improving the accuracy and robustness of ultra-short-term photovoltaic power prediction.

[0030] In the model predictive control stage, the power prediction sequence output by the iTransformer-based photovoltaic power prediction stage is used as the disturbance term. By establishing a mathematical model and constraints for the energy storage system, the power optimization allocation of the energy storage system and the smooth control of the photovoltaic grid-connected power are achieved. The specific implementation process includes four steps: control mode design, state space equation establishment, objective function design, and constraint setting. 1. Establishment of state-space equations: Define the system state variables as energy storage charged states. The controlled variable is the energy storage charging and discharging power. Combined with predicted photovoltaic power A discrete state-space model describing the energy storage charging and discharging characteristics and grid-connected power balance is established: in, To control the cycle, This refers to the rated capacity of the energy storage.

[0031] 2. Objective Function and Constraint Design: To achieve a trade-off between power smoothing and energy storage lifetime protection, this embodiment adopts a constraint-based deviation minimization strategy. Objective Function: The objective is to minimize the sum of squares of the differences between the grid-connected power and the predicted photovoltaic power, aiming to ensure the energy storage system operates only when necessary. Indicates the prediction of the first time domain i The grid-connected power prediction value is calculated in steps. It is the result of the current photovoltaic output and energy storage command. To represent the first term in the prediction time domain i Predicted photovoltaic output power in steps, N For the current predicted time, i This indicates the step index.

[0032] Constraints: Constraints include power fluctuation constraints at the grid connection point and safety constraints of the energy storage system; (1) Grid connection point power fluctuation constraint, requiring the grid connection power range in the prediction time domain to satisfy: in To allow for fluctuation coefficients, This refers to the rated installed capacity. (2) Energy storage system safety constraints include energy storage battery state of charge constraints and energy storage power constraints, which limit the upper and lower limits of energy storage charging and discharging power and the safe range of SOC: , These represent the lower and upper limits of the permissible operating state of charge (SOC) of the energy storage system, respectively. , These are the maximum charging power and maximum discharging power of the energy storage system, respectively.

[0033] A second aspect of the present invention provides a controller for smoothing photovoltaic power fluctuations. The controller includes a communication interface, a memory, and a processor. The communication interface is configured to establish bidirectional communication with the photovoltaic power plant and the hybrid energy storage system, and to establish a bidirectional communication link with the photovoltaic power plant monitoring terminal and the hybrid energy storage converter. It is capable of real-time acquisition of photovoltaic power sampling values, meteorological monitoring data, and operational data such as the energy storage system's state of charge feedback, and sends the generated power dispatch control commands to the hybrid energy storage converter. The memory is configured to store a historical time-series database, pre-trained hybrid prediction model parameters, a computer-executable control algorithm program, and a hybrid prediction model pre-trained using the method described in the first aspect of the present invention. The processor is connected to both the communication interface and the memory, and is configured to execute the computer program to perform the following operations: constructing and normalizing physical features from the acquired data; generating a photovoltaic power prediction sequence containing physical priors using the hybrid prediction model; performing rolling time-domain optimization based on the prediction sequence and the current energy storage state to generate an optimal charging and discharging power command that satisfies grid connection volatility constraints, and sending it through the communication interface.

[0034] A third aspect of the present invention provides a photovoltaic power generation grid-connected system, comprising a photovoltaic power generation unit, a hybrid energy storage system, and the aforementioned controller. The photovoltaic power generation unit converts solar energy into electrical energy and connects it to a DC bus for transmission to the grid connection point. The hybrid energy storage system includes power-type energy storage units (such as supercapacitors or flywheels) and energy-type energy storage units (such as lithium batteries) connected in parallel, used to respond to external commands to perform power throughput to regulate the power balance of the bus. The controller establishes communication connections with both the photovoltaic power generation unit and the hybrid energy storage system via communication interfaces, executes the model predictive control method based on real-time collected data, and coordinates the actions of the hybrid energy storage system to smooth the output power of the photovoltaic power generation unit.

[0035] Compared with the prior art, the present invention has the following beneficial effects: It solves the problems of prediction distortion caused by data nonstationarity and variables: by introducing VMD decomposition and RevIN reversible normalization, it effectively isolates high-frequency noise and overcomes seasonal data distribution drift; by using iTransformer's variable dimension attention mechanism, it can capture the complex nonlinear relationship between "temperature-irradiance-power", which significantly improves the prediction accuracy under complex meteorological conditions.

[0036] The physical prior guidance mechanism enhances the robustness of the model: This invention does not rely solely on data-driven approaches, but rather deeply embeds the solar trajectory (elevation angle / azimuth angle) into the attention layer of the Transformer. This physical mechanism serves as strong prior information, guiding the model to maintain a prediction trend that conforms to physical laws even when cloud cover or sudden changes in radiation occur, and fundamentally eliminating predictions that violate physical common sense (such as high power at night).

[0037] A specific MPC strategy significantly extends energy storage lifespan: This invention employs a control strategy of "volatility constraint + deviation minimization objective." Unlike traditional control methods that attempt to "forcibly straighten" the power curve, this invention's strategy aims to maximize the approximation of the original photovoltaic output while meeting the grid volatility assessment threshold. This strategy avoids frequent charging and discharging of the energy storage system during unnecessary minor fluctuations, thereby effectively reducing battery cycle losses and extending system lifespan.

[0038] In summary, this invention effectively addresses the problems of lag, insufficient prediction accuracy, and unreasonable energy storage power allocation in traditional photovoltaic power smoothing algorithms by introducing a hybrid energy storage control method that combines an iTransformer-based ultra-short-term photovoltaic power prediction model with model predictive control (MPC). Compared to power smoothing methods based solely on filtering or fixed time constants, this invention utilizes variational mode decomposition (VMD) and reversible instance normalization (RevIN) to extract multi-modal features and align the distribution of power signals, significantly improving prediction accuracy and stability. By using the iTransformer network to model multivariate time-series dependencies, it can anticipate short-term fluctuations in photovoltaic power. Combined with a rolling time-domain model predictive control strategy, it can dynamically adjust charging and discharging power based on the energy storage's state of charge, forming a closed-loop optimization mechanism of prediction and control. This method not only significantly improves the smoothness of photovoltaic system output power and control real-time performance but also considers the energy utilization efficiency and operational lifespan of the energy storage system, demonstrating good engineering feasibility and application value. Attached Figure Description

[0039] Figure 1 This is the overall technical block diagram of the present invention.

[0040] Figure 2 This is a flowchart of the data preprocessing process of the present invention.

[0041] Figure 3 This is a diagram of the internal structure of the hybrid prediction model in this invention.

[0042] Figure 4 This is a structural diagram of the model predictive control method of the present invention.

[0043] Figure 5 This is a hardware architecture diagram of the control system provided in an embodiment of the present invention. Detailed Implementation

[0044] To better understand the present invention, the hybrid energy storage predictive control method for photovoltaic power fluctuations will be further described below in conjunction with the accompanying drawings and embodiments.

[0045] This embodiment provides a hybrid energy storage predictive control method for photovoltaic power fluctuations. This method operates within the energy management system or energy storage converter controller of a photovoltaic power plant. The method includes: collecting historical and meteorological data from the photovoltaic power plant, cleaning and normalizing them to construct a multivariate time-series dataset; processing non-stationary photovoltaic power sequences using variational mode decomposition and the Reversible Instance Criterion (RevIN); constructing a prediction model based on TCN-iTransformer, extracting local fluctuation features of modal components using the TCN module, and capturing global dependencies between multiple variables through an inverted attention mechanism; introducing the solar trajectory as a physical prior, and combining cross-attention and physical gating strategies to ensure the prediction results conform to physical laws. Finally, a hybrid energy storage state-space model is established, and the high-precision prediction sequence is introduced into the MPC rolling optimization framework to solve for the optimal control command under grid connection volatility constraints, achieving effective mitigation of photovoltaic fluctuations.

[0046] like Figure 1 As shown, the three core components of this method are data preprocessing, ultra-short-term power prediction, and model predictive control, which will be explained below.

[0047] Step 1: Data Preprocessing In this embodiment, data preprocessing is the foundation for high-quality prediction. For example... Figure 2 Mainly includes: Data Acquisition: First, historical photovoltaic power data of the photovoltaic power station is acquired through the power station's data acquisition and monitoring system, along with synchronous meteorological data such as solar irradiance, ambient temperature, and solar panel temperature. The sampling interval is set to 5 minutes or 15 minutes.

[0048] Outlier Handling and Physical Constraint Verification: Physical and logical verification is performed on the collected data. For example, using prior knowledge of the solar altitude angle, if the photovoltaic power is not zero during nighttime, it is determined to be sensor drift, and the power is forcibly corrected to zero. For outliers during non-zero periods, statistical methods (such as 3) are used. Identification is based on principles.

[0049] Missing value imputation: For missing values ​​caused by data transmission interruptions, this embodiment uses the K-Nearest Neighbors (KNN) algorithm for imputation. Specifically, it finds the N most similar samples in the historical database based on the Euclidean distance formula. The weighted average of the found neighbor samples is then calculated as the target value to be imputed. Normalization: To eliminate the dimensional differences between power (kW level) and meteorological data (such as temperature °C), the RevIN normalization method was used to uniformly map the multivariate data to the [0, 1] interval. This provides a stable initial data distribution for subsequent models and constructs a standardized multivariate time-series input feature set.

[0050] Step 2: Ultra-short-term photovoltaic power prediction (corresponding to...) Figure 3 ) This step employs a hybrid prediction model based on VMD-RevIN-TCN-iTransformer, and the specific processing flow is as follows: 1. Multiscale signal decomposition (VMD) Considering that the photovoltaic power sequence contains fluctuations at different frequencies, the VMD algorithm is used to decompose the non-stationary photovoltaic power sequence into... The original sequence contains several intrinsic mode components (IMFs) with specific center frequencies. By iteratively searching for the optimal solution of the variational mode, the high-frequency random noise (such as cloud cover) in the original sequence is physically decoupled from the low-frequency diurnal variation trend (apparent solar motion), thus reducing the signal complexity.

[0051] 2. Data Distribution Alignment (RevIN) To address the statistical distribution shift problem caused by seasonal and weather variations in photovoltaic data, this embodiment introduces a Reversible Instance Normalization (RevIN) module. This module employs a symmetrical structure of "normalization—modeling—denormalization": Forward normalization: Before the data is input into the neural network, the mean and standard deviation of each IMF component and meteorological variable within the current time window are calculated, and the data is mapped to a standard normal distribution to remove non-stationary statistical features. Inverse transformation: After the model outputs the prediction results, the previously stored statistics are used to perform an inverse transformation to restore the original physical dimensions of the data.

[0052] This mechanism ensures that the model can focus on the fluctuation pattern of the learning sequence without being affected by the drift of the absolute numerical range of the data.

[0053] 3. Local Feature Extraction (TCN) The RevIN-normalized components are input as independent channels into a temporal convolutional network (TCN). TCN uses one-dimensional dilated convolutions to increase the receptive field, accurately captures local short-term abrupt changes in the power sequence, and prevents gradient vanishing through residual connections.

[0054] 4. Invert Transformer global modeling This paper introduces the iTransformer model for time series forecasting. Unlike traditional methods, this embodiment maps the complete time series of each physical variable (meteorological variable and IMF component) into a high-dimensional independent feature vector. A multi-head self-attention mechanism is used to calculate the correlation between feature vectors, thereby capturing the global nonlinear relationships between different variables.

[0055] 5. Physically Guided Cross-Attention Mechanism To strengthen the constraints of physical laws, this embodiment introduces a solar trajectory guidance branch.

[0056] Physical feature mapping: Based on the latitude, longitude and time of the predicted moment, calculate the future solar altitude angle and azimuth angle sequence, and map them into a physical prior feature vector.

[0057] Dynamic attention guidance: During the decoding phase, the physical prior feature vector is used as the query vector, and the historical variable features extracted by iTransformer are used as the key vector and value vector.

[0058] This mechanism enables the model to "dynamically adjust" the weights of different meteorological variables based on the future position of the sun, thereby generating a predictive sequence that conforms to physical logic.

[0059] 6. After the physical gated output model outputs the prediction sequence, it again uses a gating unit based on the solar altitude angle to force the predicted value at nighttime to be set to zero, and outputs a high-precision power prediction sequence for the future preset time window.

[0060] Step 3: Hybrid energy storage smoothing stage based on model predictive control (corresponding to...) Figure 4 This step, based on the prediction results, optimizes the power allocation of the hybrid energy storage system through rolling time-domain optimization. First, the system acquires the current photovoltaic output power and energy storage state of charge as the initial state for control. Then, based on the predicted photovoltaic power sequence for the next N steps, the system state in the predicted time domain is calculated, and a state-space model is established. Secondly, an optimization objective function is constructed: the combined objectives are to minimize grid-connected power fluctuation rate and minimize the operational amplitude of the energy storage system. The objective function is designed as follows: Then, the following constraints are set: During the solution process, the following physical constraints must be satisfied: Volatility constraint: The rate of change of grid-connected power must not exceed the threshold specified by the grid dispatching authority. Safe operation constraints include upper and lower limits of State of Charge (SOC) and upper and lower limits of charge and discharge power. Finally, rolling optimization and feedback correction are performed: In each control cycle, based on the latest state feedback and prediction sequence, the above optimization problem is solved to obtain the optimal control sequence for the next N steps. Control commands are then issued to drive the hybrid energy storage system. At the next moment, the system state is updated based on the sensor-measured State of Charge (SOC), and the above process is repeated. This closed-loop rolling mechanism effectively compensates for prediction errors and external environmental disturbances.

[0061] Through the above implementation methods, the present invention can utilize a high-precision physical guidance prediction model, combined with the rolling optimization capability of MPC, to achieve smooth output of photovoltaic grid-connected power and effectively protect the lifespan of energy storage components.

[0062] This embodiment provides a hardware implementation system based on the above control method. For example... Figure 5 As shown, the system mainly consists of a photovoltaic power generation array, a hybrid energy storage device, and a core controller.

[0063] 1. Core Controller Architecture The controller uses an embedded industrial control computer or a high-performance DSP platform, and its internal hardware logic includes: Communication interface unit: integrates RS485 / CAN bus and Ethernet interface. On one hand, it reads real-time photovoltaic power, irradiance, temperature and other meteorological data from the photovoltaic power plant's data acquisition unit (DAQ) via Modbus / TCP protocol; on the other hand, it communicates with the energy storage converter (PCS) via CAN bus to read the battery state of charge (SOC) and issue active power reference commands in millisecond cycles.

[0064] Storage Unit: Divided into program storage area and data storage area. The program storage area contains the VMD signal decomposition algorithm, iTransformer prediction network weights, and MPC quadratic programming solver. The data storage area maintains a circular buffer to store historical running data from the most recent 24 hours for use in sliding window prediction.

[0065] The computational processing unit (CFG) serves as the core of the algorithm execution, calling the program in the memory area via the internal bus. Its operational logic is as follows: First, it cleans and normalizes the raw data input from the interface unit; then, it runs the hybrid prediction model, calculates the solar altitude angle using the built-in clock, and outputs the power prediction curve for the next N steps; finally, it substitutes this curve as a disturbance quantity into the MPC state equation to calculate the optimal energy storage action sequence that satisfies the volatility constraint, and sends the first control quantity to the PCS for execution via the communication interface.

[0066] 2. System-level connectivity Photovoltaic power generation units: connected to the AC bus via DC / AC inverter (or connected to the DC bus via DC / DC), are the main source of power fluctuations.

[0067] Hybrid energy storage system: Composed of a high-power-density supercapacitor bank and a high-energy-density lithium iron phosphate battery bank connected in parallel. Based on calculations, the controller distributes high-frequency fluctuation commands to the supercapacitors and low-frequency smoothing commands to the lithium batteries. This ensures the smoothness of power at the grid connection point while avoiding frequent high-frequency oscillations in the lithium batteries, thereby extending the system's lifespan.

[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A model-based predictive control method for smoothing photovoltaic power fluctuations, characterized in that, Includes the following steps: Step 1: Obtain historical power sequence and meteorological correlation data of photovoltaic power station, and calculate physical parameters of solar trajectory based on geographical information of the prediction period, and construct a multivariate time series input set containing physical prior features; Step 2 involves processing the input set using a multi-level feature extraction architecture, including: decomposing the non-stationary photovoltaic power sequence into modal components of different frequencies using a multi-scale signal decomposition algorithm; capturing the local short-term dependency features of each modality and meteorological variable through a local feature extraction module; mapping each power modal component and meteorological variable into independent variable feature vectors, establishing cross-variable dependencies in the variable dimension; dynamically adjusting the model's focus weight on historical features using the physical parameters of the solar trajectory through a physical prior guidance mechanism; and correcting the initial output of the model using a gating strategy based on physical thresholds to generate a photovoltaic power prediction sequence. Step 3: Establish a state-space model that includes the dynamic characteristics of the hybrid energy storage system, and introduce the photovoltaic power prediction sequence as a feedforward disturbance into the rolling optimization framework of the model predictive control; construct the optimization objective function, set constraints, solve for the optimal charging and discharging power command at the current moment, and send the command to the converter of the hybrid energy storage system through the communication interface to drive the hybrid energy storage system to operate, so as to smooth the grid-connected power.

2. The photovoltaic power fluctuation model prediction and control method according to claim 1, characterized in that, In step 2, the multi-scale signal decomposition algorithm employs variational mode decomposition. The processing includes: decomposing the photovoltaic power sequence into K intrinsic mode components with finite bandwidth at center frequency using variational mode decomposition; employing a reversible instance normalization strategy to align the distribution of each component before feature extraction and performing an inverse transformation to restore physical dimensions before output; the local feature extraction module uses a temporal convolutional network to extract the local temporal features of each normalized component as the local short-term dependency features using an dilated causal convolutional structure.

3. The photovoltaic power fluctuation model prediction and control method according to claim 1, characterized in that, In step 2, the prediction model uses a Transformer network based on inverted dimensions, and its construction and processing include: Each meteorological variable and power mode component is mapped into a semantically independent high-dimensional variable feature vector to achieve variable embedding; The solar altitude angle and azimuth angle are calculated based on the latitude, longitude and timestamp of the photovoltaic power station, and mapped to a physical prior feature vector aligned with the variable dimension to achieve physical prior mapping; The physical prior feature vector is used as the query vector, and the historical variable features are used as the key vector and value vector. The model’s attention weight for different meteorological variables is dynamically adjusted through cross-attention operation.

4. The photovoltaic power fluctuation model prediction and control method according to claim 1, characterized in that, In step 2, The step of using a physical threshold-based gating strategy to correct the initial output of the model involves detecting the solar elevation angle at the prediction time after the model outputs the initial prediction sequence. If the solar elevation angle is lower than a preset radiation threshold, the gating logic is forcibly triggered to correct the photovoltaic power prediction value at that time to zero, and the corrected sequence is transmitted to step 3.

5. The photovoltaic power fluctuation model prediction and control method according to claim 1, characterized in that, In step 3, the state-space model is constructed as follows: in, It is an energy storage system The state of charge of the energy storage system at all times. It is an energy storage system Charge and discharge power, yes Grid-connected power at any given time It is a prediction Photovoltaic power at any given time, where T is the control period. It is the rated capacity of the energy storage system.

6. The method for predictive control of smooth photovoltaic power fluctuations according to claim 1, characterized in that, In step 3, an optimization objective function is constructed with the goal of minimizing the sum of squares of the differences between the grid-connected power and the predicted photovoltaic power, expressed as: Indicates the prediction of the first time domain i The grid-connected power prediction value with a step size is calculated jointly from the current photovoltaic output and energy storage commands. To represent the first term in the prediction time domain i Predicted photovoltaic output power in steps, N For the current predicted time, i This indicates the step index.

7. The method for predictive control of smooth photovoltaic power fluctuations according to claim 1 or 6, characterized in that, In step 3, the constraints include grid connection point power volatility constraints and energy storage system safety constraints; the grid connection point power volatility constraints require that the grid connection power range within the prediction time domain satisfy the following: in To allow for fluctuation coefficients, This refers to the rated installed capacity. The energy storage system safety constraints include energy storage battery state of charge constraints and energy storage power constraints: , These represent the lower and upper limits of the permissible operating state of charge of the energy storage system's batteries, respectively. , These are the maximum charging power and maximum discharging power of the energy storage system, respectively.

8. A controller for smoothing photovoltaic power fluctuations, characterized in that, include: The communication interface is configured to establish bidirectional communication with photovoltaic power plants and hybrid energy storage systems, collect operational data in real time, and send control commands. A memory configured to store a computer program and a hybrid prediction model pre-trained using the method of claim 1; The processor, connected to both the communication interface and the memory, is configured to execute the computer program to perform the following operations: constructing and normalizing physical features of the collected data; generating a photovoltaic power prediction sequence containing physical priors using the hybrid prediction model; performing rolling time-domain optimization based on the prediction sequence and the current energy storage state to generate an optimal charging and discharging power command that satisfies the grid connection volatility constraint, and issuing it through the communication interface.

9. A photovoltaic power generation grid-connected system, characterized in that, include: A photovoltaic power generation unit is configured to convert solar energy into electrical energy and connect it to a DC bus; a hybrid energy storage system includes a power-type energy storage unit and an energy-type energy storage unit connected in parallel, configured to adjust the power balance of the bus in response to a power command; a controller as described in claim 8 establishes communication connections with the photovoltaic power generation unit and the hybrid energy storage system respectively, and is used to execute the method as described in any one of claims 1 to 7 to generate the power command.

Images