Distributed photovoltaic energy storage power generation system and control method

Abstract

This invention relates to the field of distributed photovoltaic energy storage control technology, specifically to a distributed photovoltaic energy storage power generation system and control method. A local edge control node uses a long short-term memory network to extract features from environmental meteorological data to generate a photovoltaic power prediction curve. A centralized coordination controller receives the prediction curve and grid dispatch instructions, and establishes a rolling optimization objective function based on a model predictive control algorithm, including power constraints and state of charge constraints, to solve and output a reference value for energy storage charging and discharging power. The energy storage converter, based on the reference value, uses a composite control strategy combining feedforward compensation and proportional-integral feedback to adjust the DC bus voltage on the energy storage side to perform charging and discharging actions. This reduces the frequency of charging and discharging actions of the energy storage system and suppresses the fluctuation amplitude of the DC bus voltage.

Description

Technical Field

[0001] This invention relates to the field of distributed photovoltaic energy storage control technology, specifically to a distributed photovoltaic energy storage power generation system and control method. Background Technology

[0002] Distributed photovoltaic (PV) energy storage power generation systems typically employ PV arrays in conjunction with energy storage converters for grid-connected operation. In existing conventional control schemes, the controller directly collects the instantaneous output power of the PV array and the voltage and current data at the grid connection point. It then compares the instantaneous power with a set power threshold or grid dispatch command, generating a control signal based on the difference. Upon receiving the control signal, the energy storage converter uses a simple proportional-integral feedback algorithm to adjust the DC bus voltage on the energy storage side, controlling the charging or discharging of the energy storage battery to smooth out fluctuations in PV power. This real-time feedback control, based on the current power deviation, relies on immediate responses to the PV output status and does not require prior prediction of future power changes.

[0003] Based on the control logic of the existing technology, photovoltaic (PV) output is highly volatile and random due to changes in meteorological conditions. Existing technologies rely solely on the current power deviation for delayed adjustments, failing to predict future PV power trends. When PV output changes drastically, the controller only activates the energy storage system after detecting a power limit exceedance. This lack of proactive control results in the energy storage system remaining in a passive, following state, leading to frequent charge-discharge switching and consequently, severe fluctuations in the DC bus voltage. Summary of the Invention

[0004] The purpose of this invention is to provide a distributed photovoltaic energy storage power generation system and control method, which can solve the problems in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A distributed photovoltaic (PV) energy storage power generation system includes a PV array, an energy storage converter, a local edge control node, and a centralized coordination controller. The local edge control node is connected to the output of the PV array to collect instantaneous PV output power and environmental meteorological data. The local edge control node uses a long short-term memory (LSTM) network to extract features from the environmental meteorological data to generate a PV power prediction curve for a preset time period. The input of the centralized coordination controller is connected to the output of the local edge control node to receive the PV power prediction curve and grid dispatch instructions. The centralized coordination controller establishes a rolling optimization objective function based on a model predictive control algorithm, which includes energy storage charging and discharging power constraints and upper and lower limits of state of charge constraints. By solving the rolling optimization objective function, it outputs a reference value for energy storage charging and discharging power in the next control cycle. The communication terminal of the energy storage converter is connected to the output of the centralized coordination controller to receive the reference value for energy storage charging and discharging power. The energy storage converter adjusts the DC bus voltage on the energy storage side according to the reference value for energy storage charging and discharging power using a composite control strategy combining feedforward compensation and proportional-integral feedback to perform charging and discharging actions.

[0006] Preferably, the local edge control node includes a meteorological data preprocessing sub-node and a prediction model inference sub-node; the meteorological data preprocessing sub-node performs time alignment and missing value imputation processing on the environmental meteorological data, and divides the processed environmental meteorological data into meteorological feature sequences according to a preset time window; the prediction model inference sub-node embeds a long short-term memory network, inputs the meteorological feature sequences into the long short-term memory network, calculates cell state update values ​​and hidden layer state vectors through the forget gate, input gate and output gate of the long short-term memory network, and maps the hidden layer state vectors to photovoltaic power prediction curves; the local edge control node packages the photovoltaic power prediction curves into data frames with timestamps and sends them to the centralized coordination controller.

[0007] Preferably, the centralized coordination controller includes an optimization model construction sub-node and a solver sub-node. The optimization model construction sub-node establishes an objective function term with the optimization objectives of minimizing grid connection point power fluctuations and minimizing the frequency of energy storage charging and discharging actions. The optimization model construction sub-node establishes inequality constraint terms using the upper limit of energy storage charging and discharging power, the lower limit of energy storage charging and discharging power, the current state of charge value, the upper limit of the state of charge, and the lower limit of the state of charge as constraints. The optimization model construction sub-node combines the objective function term and the inequality constraint terms to generate a rolling optimization objective function. The solver sub-node calls the interior point method to iteratively optimize the rolling optimization objective function and outputs the energy storage charging and discharging power sequence within the future preset control time domain. The solver sub-node extracts the first value in the energy storage charging and discharging power sequence as the energy storage charging and discharging power benchmark value for the next control cycle.

[0008] Preferably, the energy storage converter includes a reference value analysis sub-node, a feedforward control loop sub-node, and a proportional-integral feedback control loop sub-node. The reference value analysis sub-node performs a lookup table conversion on the energy storage charging and discharging power reference value to obtain the energy storage side DC current reference value. The feedforward control loop sub-node collects the energy storage side DC bus voltage and energy storage side DC current, calculates the current instantaneous power based on the energy storage side DC bus voltage and energy storage side DC current, and uses the difference between the current instantaneous power and the energy storage charging and discharging power reference value as the feedforward disturbance compensation amount input to the feedforward control loop sub-node. The proportional-integral feedback control loop sub-node uses the difference between the energy storage side DC current reference value and the energy storage side DC current as the input of the proportional-integral regulator, and the output of the proportional-integral regulator is superimposed with the feedforward disturbance compensation amount to generate a pulse width modulation signal to drive the power switching transistors of the energy storage converter.

[0009] Preferably, a communication link based on the User Datagram Protocol (UDP) is established between the local edge control node and the centralized coordination controller. The local edge control node adds a frame type identifier and a sequence number to the frame header of the data frame and a cyclic redundancy check (CRC) code to the frame tail. After receiving the data frame, the centralized coordination controller parses the frame type identifier. When the frame type identifier indicates a photovoltaic power prediction curve, the centralized coordination controller extracts the sequence number and performs time-series rearrangement to restore the time sequence of the photovoltaic power prediction curve. A communication link based on the controller local area network (Controller Area Network) is established between the centralized coordination controller and the energy storage converter. The centralized coordination controller sends Controller Area Network messages containing reference values ​​for energy storage charging and discharging power to the energy storage converter at preset equal time intervals.

[0010] Preferably, the centralized coordination controller has a built-in communication watchdog timer and a degraded control trigger sub-node. When the communication watchdog timer detects that the local edge control node has continuously lost a preset number of photovoltaic power prediction curve data frames, the communication watchdog timer outputs an abnormal trigger signal to the degraded control trigger sub-node. After receiving the abnormal trigger signal, the degraded control trigger sub-node masks the prediction correlation term in the rolling optimization objective function, switches the optimization objective function to a constant power control mode that tracks the grid dispatch command, and calculates the degraded energy storage charging and discharging power reference value based on the current energy storage state of charge value and the grid dispatch command, and sends the degraded energy storage charging and discharging power reference value to the energy storage converter.

[0011] Preferably, the prediction model inference sub-node is connected to an online model update sub-node, which stores historical meteorological feature sequences and historical actual photovoltaic output power. The online model update sub-node, according to a preset training period, inputs the historical meteorological feature sequences into a long short-term memory network to obtain historical predicted power, and calculates the mean squared error loss value between the historical predicted power and the historical actual photovoltaic output power. The online model update sub-node uses a gradient descent algorithm to calculate the gradient of the weight matrix of each network layer in the long short-term memory network based on the mean squared error loss value, and updates the weight matrix using a backpropagation mechanism. The online model update sub-node then synchronizes the updated weight matrix to the prediction model inference sub-node.

[0012] Preferably, the optimization model construction sub-node is further connected to a dynamic state of charge boundary calculation sub-node; the dynamic state of charge boundary calculation sub-node collects ambient temperature data and cycle count data of the energy storage battery in real time, queries a preset temperature-internal resistance mapping table based on the ambient temperature data to obtain the current battery internal resistance value, and calculates the maximum usable capacity decay coefficient of the energy storage battery by combining the current battery internal resistance value and cycle count data; the optimization model construction sub-node corrects the initial upper limit and lower limit of the initial state of charge downward based on the maximum usable capacity decay coefficient, and uses the corrected upper limit and lower limit of the state of charge as the upper and lower limit constraints of the state of charge in the rolling optimization objective function.

[0013] Preferably, the proportional-integral feedback control loop sub-node has embedded anti-integral saturation processing logic. When the output of the proportional-integral regulator reaches a preset output limit value, the anti-integral saturation processing logic freezes the integral accumulation operation inside the proportional-integral regulator and calculates the product of the difference between the energy storage side DC current reference value and the energy storage side DC current and the proportional gain. When the product is less than the preset output limit value and the polarity of the difference is reversed, the anti-integral saturation processing logic unfreezes and restarts the integral accumulation operation according to the current difference value. The proportional-integral feedback control loop sub-node generates the output of the proportional-integral regulator based on the restarted integral accumulation result.

[0014] A distributed photovoltaic energy storage power generation control method includes: Step 1, collecting instantaneous photovoltaic output power and environmental meteorological data of the photovoltaic power generation array through a local edge control node, and extracting features from the environmental meteorological data using a long short-term memory network to generate a photovoltaic power prediction curve within a preset time period; Step 2, receiving the photovoltaic power prediction curve and grid dispatch instructions through a centralized coordination controller, establishing a rolling optimization objective function based on a model predictive control algorithm that includes energy storage charging and discharging power constraints and upper and lower limits of state of charge constraints, and outputting the energy storage charging and discharging power reference value for the next control cycle by solving the rolling optimization objective function; Step 3, receiving the energy storage charging and discharging power reference value through an energy storage converter, and adjusting the DC bus voltage on the energy storage side according to the energy storage charging and discharging power reference value using a composite control strategy combining feedforward compensation and proportional-integral feedback to perform charging and discharging actions.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention utilizes a local edge control node and a long short-term memory network to extract features from environmental meteorological data to generate a photovoltaic power prediction curve. A centralized coordination controller then uses a model predictive control algorithm to establish a rolling optimization objective function incorporating power constraints and state-of-charge constraints to derive a reference value for energy storage charging and discharging power. This allows the control system to plan energy storage actions in advance based on future power variation trends. Combined with a composite control strategy using feedforward compensation and proportional-integral feedback, the energy storage converter adjusts the DC bus voltage based on the power reference value. This directly converts the data-driven prediction results into control signals for the underlying power electronic converter, overcoming the lag inherent in existing technologies that rely solely on real-time feedback. This reduces the frequency of charging and discharging actions in the energy storage system and suppresses fluctuations in the DC bus voltage.

[0016] 2. This invention enables the power prediction model to adapt to changes in the operating status of photovoltaic modules by setting up online update sub-nodes at the local edge control node, calculating the mean square error loss value based on historical data, and updating the weight matrix of the long short-term memory network. It also uses a dynamic state-of-charge boundary calculation sub-node to downwardly correct the upper and lower limits of the state of charge constraints based on ambient temperature and cycle count, preventing the use of energy storage power beyond the physical boundaries of the battery. Furthermore, by embedding anti-integral saturation processing logic within the proportional-integral feedback control loop sub-nodes, it freezes the integral accumulation when the output reaches the limit and restarts it when the difference polarity reverses, preventing output runaway during the converter's adjustment process and ensuring the system's operational stability under extreme conditions. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall control process of the distributed photovoltaic energy storage power generation system of the present invention. Figure 2This is a flowchart of the local edge control node meteorological data preprocessing and photovoltaic power prediction process of the present invention; Figure 3 This is a flowchart of the long short-term memory network gating computation and cell state update process of the present invention; Figure 4 This is a flowchart illustrating the construction and solution of the rolling optimization objective function for the centralized coordination controller of the present invention. Figure 5 This is a flowchart of the dynamic boundary correction and degradation control triggering process for the state of charge of the present invention. Figure 6 This is a flowchart of the feedforward-proportional-integral composite control and anti-integral saturation processing of the energy storage converter of the present invention. Detailed Implementation

[0018] This specific embodiment relates to the field of distributed photovoltaic energy storage power generation technology. The disclosed technical solution is designed for the grid-connected operation control scenario of a distributed photovoltaic energy storage power generation system. The following provides a complete and clear description of the technical solution in conjunction with a specific implementation scenario. Those skilled in the art can fully reproduce the corresponding technical solution and solve the corresponding technical problems in the background art based on the content disclosed in this embodiment.

[0019] refer to Figure 1 In one embodiment, the distributed photovoltaic energy storage power generation system includes a photovoltaic power generation array, an energy storage converter, a local edge control node, and a centralized coordination controller. The photovoltaic power generation array consists of multiple sets of photovoltaic modules connected in series and parallel. Its output is connected to the analog signal acquisition port of the local edge control node via DC voltage and current transmitters. The DC output of the photovoltaic power generation array is simultaneously connected to a DC bus, which is connected to the DC side port of the energy storage converter. The AC side port of the energy storage converter is connected to the grid connection point. The Ethernet communication port of the local edge control node establishes a communication connection with the WAN communication port of the centralized coordination controller, and the fieldbus communication port of the centralized coordination controller establishes a communication connection with the communication terminal of the energy storage converter.

[0020] The local edge control node acquires the DC voltage and DC current at the output of the photovoltaic array in real time through its analog acquisition port, and calculates the instantaneous output power of the photovoltaic array based on the acquired DC voltage and DC current. Simultaneously, the local edge control node collects environmental meteorological data at the corresponding locations through connected irradiance sensors, ambient temperature sensors, wind speed sensors, wind direction sensors, and relative humidity sensors. The sampling frequency of the photovoltaic instantaneous output power is consistent with the sampling frequency of the environmental meteorological data. The local edge control node uses a long short-term memory network to extract time-series features from the acquired environmental meteorological data, generating a photovoltaic power prediction curve for a preset duration. The preset duration matches the time period division rules of the power grid dispatch, and the time step of the prediction curve is consistent with the duration of subsequent control cycles.

[0021] The centralized coordination controller receives photovoltaic power prediction curves from local edge control nodes via a communication link at its input end. Simultaneously, it receives grid dispatch instructions from the grid dispatching agency via a dispatch communication port. These instructions include the planned power curve at the grid connection point, peak-valley scheduling constraints, and power factor operation requirements. Based on a model predictive control algorithm, the centralized coordination controller establishes a rolling optimization objective function that includes constraints on energy storage charging and discharging power and upper and lower limits of state of charge. The photovoltaic power prediction curves and grid dispatch instructions are used as input boundary conditions for the rolling optimization objective function. By solving the rolling optimization objective function, the controller outputs the baseline value of energy storage charging and discharging power for the next control cycle.

[0022] The communication terminal of the energy storage converter receives the energy storage charging and discharging power reference value output by the centralized coordination controller through the fieldbus communication link. Based on the received energy storage charging and discharging power reference value, the energy storage converter adopts a composite control strategy combining feedforward compensation and proportional-integral feedback to adjust the DC bus voltage on the energy storage side and drive the internal power switching tubes to perform the corresponding charging or discharging actions.

[0023] Specifically, in this embodiment, the time step for photovoltaic power prediction, the control cycle for rolling optimization, and the time interval for data sampling all adopt a unified timing reference to avoid control deviations caused by timing misalignments in different stages. The corresponding timing parameters in this embodiment are shown in the table below.

[0024] Table 1. Correspondence between timing parameters for photovoltaic power prediction and rolling optimization control

[0025] In this embodiment, the output power acquisition of the photovoltaic power generation array and the acquisition of environmental meteorological data are completed at the local edge, avoiding the remote transmission of massive amounts of raw sampling data. The photovoltaic power prediction based on the long short-term memory network is completed at the local edge control node, and only the prediction results are uploaded to the centralized coordination controller, reducing the bandwidth occupation of the communication link. The centralized coordination controller completes rolling optimization under multiple constraints based on the prediction results and scheduling instructions, and outputs the control reference value. The energy storage converter completes the bottom-level execution based on the feedforward and feedback composite control strategy, forming a three-level control architecture of local prediction-centralized optimization-bottom-level execution. This realizes the forward planning of energy storage charging and discharging actions, overcomes the defects of existing technologies that rely solely on real-time power deviation for lag adjustment, reduces the charging and discharging switching frequency of the energy storage system, and suppresses the fluctuation amplitude of DC bus voltage.

[0026] refer to Figure 2In a preferred embodiment, the local edge control node includes a meteorological data preprocessing sub-node and a prediction model inference sub-node. The meteorological data preprocessing sub-node receives collected environmental meteorological data and photovoltaic instantaneous output power data. First, it performs time alignment processing on the multi-source collected data. Using the sampling timestamp of the photovoltaic instantaneous output power as a reference, it performs linear interpolation alignment on the environmental meteorological data collected by each meteorological sensor, ensuring that the time reference of all collected data at the same timestamp is consistent and eliminating temporal misalignment caused by sampling time deviations of different sensors. The meteorological data preprocessing sub-node then performs missing value imputation processing on the time-aligned environmental meteorological data. For meteorological data with fewer than a preset threshold of consecutive missing values, it uses linear interpolation of adjacent sampling points to complete the imputation; for meteorological data with more than or equal to the preset threshold of consecutive missing values, it uses the average of historical meteorological data for the same time period to complete the imputation, avoiding interference from missing data on subsequent feature extraction and prediction results. The meteorological data preprocessing sub-node will complete the preprocessing of environmental meteorological data and divide it into meteorological feature sequences according to a preset time window. Each time step's meteorological feature sequence contains five features: irradiance, ambient temperature, wind speed, wind direction, and relative humidity at the corresponding timestamp. These features form a two-dimensional feature sequence with dimensions of [time step, feature dimension], which is then output to the prediction model inference sub-node.

[0027] refer to Figure 3 The prediction model's inference sub-node embeds a Long Short-Term Memory (LSTM) network. The LSM network structure includes an input layer, two cascaded LSTM layers, a fully connected layer, and an output layer. The input layer receives the meteorological feature sequence output by the meteorological data preprocessing sub-node, and the output layer outputs the photovoltaic power prediction value for the corresponding time step. The prediction model's inference sub-node inputs the meteorological feature sequence into the LSM network and calculates the cell state update value and hidden layer state vector through the LSM network's forget gate, input gate, and output gate. The specific calculation process is as follows.

[0028] The forgetting gate controls the degree to which information from the cell's previous state is retained or forgotten. The calculation formula is as follows: ; in, Let be the output vector of the forget gate at time t. It is the sigmoid activation function. Here is the weight matrix for the forget gate. Let be the hidden layer state vector at time t-1. Let be the meteorological feature vector input at time t. Let be the bias vector of the forget gate. This is a concatenation operation between two vectors.

[0029] The input gate controls the degree to which the feature information input at the current time is written into the cell state. It includes two parts: input gating and candidate cell state calculation. The calculation formula is as follows: ; ; in, Let be the output vector of the input gate at time t. Here is the weight matrix of the input gate. The bias vector for the input gate; Let be the candidate cell state vector at time t. This is the weight matrix for the candidate cell states. This is the bias vector for the candidate cell state. It is the hyperbolic tangent activation function.

[0030] Based on the calculation results of the forget gate and the input gate, the cell state is updated. The calculation formula is as follows: ; in, Let be the cell state vector updated at time t. Let be the cell state vector at time t-1. This is the Hadamard product operation, which is the product operation of corresponding elements of a vector.

[0031] The output gate controls the extent to which information from the cell state is output to the hidden layer state, and also controls the calculation of the hidden layer state vector. The calculation formula is as follows: ; ; in, Let be the output vector of the output gate at time t. This is the weight matrix of the output gate. This is the bias vector for the output gate; Let t be the hidden layer state vector at time t.

[0032] The prediction model inference sub-node inputs the hidden layer state vector output from the Long Short-Term Memory network into the fully connected layer. Through linear mapping and activation function processing in the fully connected layer, the hidden layer state vector is mapped to the photovoltaic power prediction value at the corresponding time step. The photovoltaic power prediction values ​​of multiple consecutive time steps constitute the photovoltaic power prediction curve within a preset time period. The local edge control node packages the generated photovoltaic power prediction curve, along with the timestamp information of the corresponding time step, into a standardized data frame and sends it to the centralized coordination controller.

[0033] A communication link based on the User Datagram Protocol (UDP) is established between the local edge control node and the centralized coordination controller. The data frames adopt a fixed frame structure, including a frame header, payload, and trailer. The frame header includes a frame type identifier and a sequence number. The frame type identifier is a fixed-length unsigned integer used to distinguish the service type of the data frame, including photovoltaic power prediction curve data frames, real-time operation data frames, and fault alarm data frames. The sequence number is a continuously incrementing unsigned integer used to identify the transmission sequence of the data frames, avoiding out-of-order transmission issues. The payload encapsulates the numerical data of the photovoltaic power prediction curve and the corresponding timestamp information. The trailer includes a 32-bit cyclic redundancy check (CRC) code, calculated based on all data in the frame header and payload, used to check for bit errors during data frame transmission. After receiving a data frame, the centralized coordination controller first performs a data integrity check based on the cyclic redundancy check code at the end of the frame. After the check passes, it parses the frame type identifier in the frame header. When the frame type identifier indicates that it is a photovoltaic power prediction curve data frame, the centralized coordination controller extracts the sequence number in the frame header, rearranges the timing of the multiple received data frames, restores the time order of the photovoltaic power prediction curve, and completes the reception and parsing of the prediction data.

[0034] The prediction model inference sub-node is connected to the online model update sub-node. The online model update sub-node is equipped with non-volatile storage media containing historical meteorological feature sequences and corresponding historical actual photovoltaic output power data. The historical data is partitioned by day and stored for at least one year. The online model update sub-node initiates the weight update process of the Long Short-Term Memory (LSTM) network according to a preset training period, which matches the preset duration of the photovoltaic power prediction curve. The online model update sub-node reads the historical meteorological feature sequences and historical actual photovoltaic output power from the storage media within the training period, inputs the historical meteorological feature sequences into the LTM network to obtain the corresponding historical predicted power, and calculates the mean squared error loss value between the historical predicted power and the historical actual photovoltaic output power. The calculation formula is as follows: ; in, This is the mean squared error loss value. The total number of training samples. The historical predicted power corresponding to the nth training sample. This represents the historical actual photovoltaic output power corresponding to the nth training sample.

[0035] The online model update sub-node employs the gradient descent algorithm. Based on the calculated mean squared error loss value, it calculates the gradients of the weight matrices and bias vectors of each network layer in the Long Short-Term Memory (LSTM) network. Gradient calculation uses a backpropagation algorithm that expands along the time dimension, calculating the partial derivatives of the loss function with respect to each weight matrix and bias vector. The online model update sub-node updates each weight matrix and bias vector according to a preset learning rate based on the calculated gradients. The learning rate uses an exponential decay strategy, initially set at 0.001, decreasing to 0.9 times its original value every 10 training epochs. After updating the weight matrices and bias vectors, the online model update sub-node synchronizes the updated network parameters to the prediction model inference sub-node, replacing the original network parameters in the prediction model inference sub-node, thus completing the online update of the LSM network.

[0036] In this embodiment, the structure and parameters of each layer of the Long Short-Term Memory network are defined as shown in the table below.

[0037] Table 2. Long Short-Term Memory Network Structure and Parameter Definitions

[0038] In this embodiment, time synchronization and missing value completion of multi-source collected data were achieved through meteorological data preprocessing, ensuring the quality of the input feature sequence. The long short-term memory network gating mechanism was used to extract long-term dependencies of time-series meteorological features, improving the accuracy of photovoltaic power prediction. Low-latency and high-reliability transmission of prediction data was achieved through a communication link based on the User Datagram Protocol and a standardized data frame structure. Adaptive parameter optimization of the long short-term memory network was achieved by updating the model's sub-nodes online, enabling the prediction model to adapt to changes in the operating status of photovoltaic modules and fluctuations in seasonal meteorological characteristics, providing a stable and accurate input basis for subsequent rolling optimization control.

[0039] refer to Figure 4 In a preferred embodiment, the centralized coordination controller includes an optimization model construction sub-node and a solver sub-node. The optimization model construction sub-node receives the photovoltaic power prediction curve and grid dispatch instructions, and establishes an objective function term with the optimization objectives of minimizing power fluctuations at the grid connection point and minimizing the frequency of energy storage charging and discharging operations. Simultaneously, the optimization model construction sub-node obtains the operating boundary parameters of the energy storage system, and establishes inequality constraint terms with the upper limit of energy storage charging and discharging power, the lower limit of energy storage charging and discharging power, the current state of charge value, the upper limit of state of charge, and the lower limit of state of charge as constraints. The optimization model construction sub-node combines the objective function term and the inequality constraint terms to generate a rolling optimization objective function.

[0040] The rolling optimization objective function is established based on a rolling time-domain optimization framework using model predictive control, with a prediction time-domain length of [missing information]. The control time domain length is And satisfy The formula for calculating the objective function of rolling optimization is: ; in, To optimize the objective function for rolling, To predict the grid connection point power at the k-th time step in the time domain, Let be the reference power of the grid connection point at the k-th time step in the power grid dispatch instruction. The weighting coefficients for the power point tracking term are: To control the change in energy storage charging and discharging power at the k-th time step in the time domain, The weighting coefficients for the frequency suppression term of energy storage charging and discharging operations.

[0041] The grid-connected power at the k-th time step in the time domain is predicted, calculated based on the photovoltaic power prediction curve and the energy storage charging and discharging power. The calculation formula is as follows: ; in, To predict the photovoltaic power at the k-th time step in the time domain, the photovoltaic power prediction curve is used. To predict the energy storage charging and discharging power at the k-th time step in the time domain, the value is positive during charging and negative during discharging.

[0042] The optimization model constructs sub-nodes that establish inequality constraints, including energy storage charging and discharging power constraints and upper and lower limits of state of charge constraints. The energy storage charging and discharging power constraint limits the charging and discharging power of the energy storage system within a single control cycle to ensure it does not exceed the rated operating range of the equipment. The constraint formula is: ; in, This represents the lower limit of the energy storage charging and discharging power, corresponding to the maximum discharge power of the energy storage system. This represents the upper limit of energy storage charging and discharging power, corresponding to the maximum charging power of the energy storage system.

[0043] State of charge (SOC) upper and lower limits are used to restrict the SOC of energy storage batteries to within a safe operating range, preventing damage from overcharging or over-discharging. First, a time-domain update model for the SOC is established, with the calculation formula as follows: ; in, This represents the energy storage state of charge at the k-th time step. This represents the energy storage state of charge at time step k-1. To improve the charging efficiency of energy storage systems. For the discharge efficiency of the energy storage system To control the duration of the cycle, This refers to the rated capacity of the energy storage battery. This is an indicator function that takes the value 1 when the condition within the parentheses is true, and 0 otherwise.

[0044] Based on the time-domain update model of the state of charge, upper and lower bound constraints on the state of charge are established, and the constraint formulas are as follows: ; in, This is the limit of the state of charge. This represents the upper limit of the state of charge.

[0045] The solver sub-node receives the rolling optimization objective function generated by the optimization model construction sub-node, and uses the interior-point method to iteratively optimize the constrained quadratic programming problem. The iteration terminates when the difference between the objective function values ​​of two adjacent iterations is less than a preset convergence threshold, or when the number of iterations reaches a preset maximum number of iterations. After solving, the solver sub-node outputs the energy storage charging and discharging power sequence in the future control time domain. The sequence length is consistent with the control time domain length. The solver sub-node extracts the first value in the energy storage charging and discharging power sequence as the benchmark value for the energy storage charging and discharging power in the next control cycle and sends it to the energy storage converter. After solving the current control cycle, the next control cycle begins. The optimization model construction sub-node re-acquires the latest photovoltaic power prediction curve and grid dispatch instructions, updates the rolling optimization objective function, and the solver sub-node repeats the above iterative optimization process to achieve rolling time-domain optimization control.

[0046] The centralized coordination controller incorporates a communication watchdog timer and a degraded control trigger sub-node. The timing period of the communication watchdog timer is consistent with the transmission period of the photovoltaic power prediction curve data frames from the local edge control node. Each time the centralized coordination controller receives a valid photovoltaic power prediction curve data frame, it performs a "feed" operation on the communication watchdog timer, resetting the timer's count value. When the communication watchdog timer detects that the local edge control node has continuously lost a preset number of photovoltaic power prediction curve data frames, it outputs an abnormal trigger signal to the degraded control trigger sub-node. Upon receiving the abnormal trigger signal, the degraded control trigger sub-node masks the prediction correlation term in the rolling optimization objective function and switches the optimization objective function to a constant power control mode that tracks grid dispatch instructions. In this mode, the optimization objective is only to track the constant power reference value at the grid connection point in the grid dispatch instructions, while the constraints remain unchanged, including the energy storage charging / discharging power constraint and the upper and lower limits of the state of charge (SOC). Based on the current SOC value and the grid dispatch instructions, the degraded energy storage charging / discharging power reference value is calculated and sent to the energy storage converter. When the centralized coordination controller receives a preset number of consecutive valid photovoltaic power prediction curve data frames, the degradation control trigger sub-nodes will deactivate the degradation operation mode and restore the normal rolling optimization control mode.

[0047] refer to Figure 5 The optimization model construction sub-node is also connected to a state-of-charge (POC) dynamic boundary calculation sub-node. This sub-node collects real-time ambient temperature data within the energy storage battery cabinet, as well as battery cycle count data uploaded by the energy storage battery management system. Based on the collected ambient temperature data, the POC dynamic boundary calculation sub-node queries a pre-defined temperature-internal resistance mapping table to obtain the current battery internal resistance value. This mapping table is generated by fitting the energy storage battery's factory test data and historical operating data, and stores the battery's DC internal resistance values ​​corresponding to different ambient temperature ranges. Combining the current battery internal resistance value with the cycle count data, the POC dynamic boundary calculation sub-node calculates the maximum usable capacity decay coefficient of the energy storage battery. The calculation formula is as follows: ; in, The maximum available capacity attenuation coefficient has a value range of (0,1]. The weighting coefficient is determined by the influence of internal resistance. The number of iterations affects the weighting coefficient; This is the current internal resistance value of the battery. This is the rated internal resistance value of the battery; This represents the current cycle count of the battery. The number of cycles is the battery's rated cycle life.

[0048] The optimization model constructs sub-nodes that, based on the calculated maximum available capacity decay coefficient, downwardly adjust the upper limit and lower limit of the initial state of charge. The adjustment formula is as follows: ; ; in, This is the corrected upper limit of the state of charge. This represents the upper limit of the initial state of charge; This is the corrected state of charge limit. This is the limit of the initial charging state.

[0049] The optimization model construction sub-node uses the modified upper and lower limits of the charge state as the upper and lower limit constraints of the charge state in the rolling optimization objective function, replacing the original fixed upper and lower limit constraints, and realizing the dynamic adjustment of the charge state constraints.

[0050] In this embodiment, the dynamic boundary correction parameters for the energy storage state of charge under different operating conditions are shown in the table below.

[0051] Table 3 Mapping Table of Dynamic Boundary Correction Parameters for Energy Storage State of Charge

[0052] In this embodiment, a multi-objective collaborative optimization of grid-connected point power tracking and energy storage charging and discharging frequency suppression is achieved through the rolling optimization objective function of model predictive control; a fast iterative solution to the constrained quadratic programming problem is achieved through the interior-point method, ensuring the real-time output of the control baseline value; a communication watchdog timer and a degradation control trigger sub-node are used to achieve degraded system operation under abnormal communication conditions, avoiding system loss of control caused by communication interruption; and a dynamic state of charge boundary calculation sub-node is used to achieve dynamic adjustment of state of charge constraints according to battery operating status and environmental conditions, avoiding power calls exceeding the physical available boundary of the battery, improving the safety of energy storage battery operation, and extending battery life.

[0053] refer to Figure 6In a preferred embodiment, the energy storage converter includes a reference value resolution sub-node, a feedforward control loop sub-node, and a proportional-integral feedback control loop sub-node. The energy storage converter adopts a bidirectional non-isolated DC / DC converter topology. The low-voltage side of the converter is connected to the energy storage battery pack, and the high-voltage side is connected to the DC bus. The charging and discharging control of the energy storage battery pack is realized by adjusting the on and off timing of the internal IGBT power switches. A communication link based on a controller area network is established between the centralized coordination controller and the energy storage converter. The communication rate is matched with the transmission distance of the fieldbus. The centralized coordination controller sends controller area network messages containing the energy storage charging and discharging power reference value to the energy storage converter at preset equal intervals. The message adopts an extended frame format. The identifier field contains the device address and message type information, and the data field encapsulates the 32-bit floating-point data of the energy storage charging and discharging power reference value.

[0054] The reference value parsing sub-node receives controller LAN messages, performs message verification and parsing, and extracts the energy storage charging and discharging power reference value. First, it limits the power reference value, ensuring the limiting range matches the rated charging and discharging power range of the energy storage converter to avoid inputting control commands exceeding the equipment's rated operating range. The reference value parsing sub-node then acquires the current terminal voltage of the energy storage battery pack. Combined with the limited energy storage charging and discharging power reference value, it calculates the reference value of the DC current on the energy storage side through a lookup table. The lookup table used is a voltage-power-current mapping table for the energy storage battery. This table is pre-established based on the charging and discharging characteristic test data of the energy storage battery and stores the DC current values ​​corresponding to different battery terminal voltages and different charging and discharging powers. The reference value parsing sub-node obtains the reference value of the DC current on the energy storage side that matches the current operating conditions through linear interpolation.

[0055] The feedforward control loop sub-nodes acquire the DC bus voltage and DC current of the energy storage side in real time using high-precision voltage and current sensors, with a sampling frequency of no less than 10kHz to ensure the real-time performance and accuracy of the sampled data. Based on the acquired DC bus voltage and DC current, the feedforward control loop sub-nodes calculate the instantaneous output power of the energy storage side. The difference between the current instantaneous power and the reference value of the energy storage charging / discharging power is used as the feedforward disturbance compensation amount and input to the feedforward control loop. The formula for calculating the feedforward disturbance compensation amount is as follows: ; in, This is the feedforward disturbance compensation amount. The feedforward gain coefficient is based on the pre-tuned main circuit parameters of the energy storage converter. This is the reference value for energy storage charging and discharging power. This represents the instantaneous output power of the current energy storage side.

[0056] The feedforward control loop sub-nodes compensate for power changes caused by photovoltaic power fluctuations and grid voltage disturbances in advance through the feedforward disturbance compensation amount, thereby overcoming the lag of the feedback control loop and suppressing the fluctuation amplitude of DC bus voltage.

[0057] The proportional-integral feedback control loop sub-node uses the difference between the reference value of the energy storage side DC current output by the reference value parsing sub-node and the DC current of the energy storage side collected by the feedforward control loop sub-node as the input of the proportional-integral regulator. The output calculation formula of the proportional-integral regulator is as follows: ; in, The output value of the proportional-integral controller. This is the proportional gain coefficient. This is the integral gain coefficient. This is the difference between the reference value and the actual collected value of the DC current on the energy storage side. For the current sampling period, The duration for accumulating the current points.

[0058] The proportional-integral feedback control loop sub-node superimposes the output value of the proportional-integral regulator with the feedforward disturbance compensation output of the feedforward control loop sub-node to generate a pulse width modulation (PWM) signal. The PWM signal is compared with a triangular carrier wave of a preset frequency to generate a PWM signal with an adjustable duty cycle. After being amplified by the drive circuit, the PWM signal drives the IGBT power switch inside the energy storage converter to regulate the DC bus voltage on the energy storage side and complete the corresponding charging or discharging action.

[0059] The proportional-integral feedback (PIF) control loop sub-nodes embed anti-integral saturation processing logic. This logic employs a limit-based integration reduction method, with preset output limits including upper and lower limits, corresponding to the modulation amplitude values ​​of the maximum and minimum duty cycles of the pulse width modulation signal. When the PIF controller's output value reaches the preset output limit, the anti-integral saturation processing logic freezes the internal integral accumulation operation of the PIF controller, stopping the integral accumulation of the current error to avoid integral saturation caused by the continuous accumulation of the integral term. Simultaneously, the anti-integral saturation processing logic calculates in real-time the product of the difference between the energy storage side DC current reference value and the energy storage side DC current and the proportional gain. When this product is less than the preset output limit and the polarity of the difference reverses, the anti-integral saturation processing logic unfreezes the integral accumulation and restarts the integral accumulation operation according to the current difference. The PIF control loop sub-nodes then generate the PIF controller's output value based on the restarted integral accumulation result.

[0060] In this embodiment, the parameter configuration of the composite control loop of the energy storage converter is shown in the table below.

[0061] Table 4. Parameter Configuration Table for Composite Control Loop of Energy Storage Converter

[0062] In this embodiment, a composite control strategy combining feedforward compensation and proportional-integral feedback is employed to achieve rapid tracking and zero steady-state error regulation of energy storage charging and discharging power. The feedforward control loop provides advance compensation for power disturbances, overcoming the lag inherent in simple proportional-integral feedback control. The proportional-integral feedback loop eliminates control deviations in a steady state. Anti-integral saturation processing logic prevents integral saturation of the proportional-integral regulator under extreme conditions, preventing output runaway during converter regulation and improving the dynamic response performance and steady-state stability of the control loop. A communication link based on the controller area network ensures high real-time and high-reliability transmission of control commands, guaranteeing the coordinated operation of the underlying actuators and the upper-level centralized control unit.

Claims

1. A distributed photovoltaic energy storage power generation system, characterized in that, This includes photovoltaic power generation arrays, energy storage converters, local edge control nodes, and centralized coordination controllers; The local edge control node is connected to the output end of the photovoltaic power generation array to collect instantaneous photovoltaic output power and environmental meteorological data. The local edge control node uses a long short-term memory network to extract features from the environmental meteorological data to generate a photovoltaic power prediction curve within a preset time period. The input terminal of the centralized coordination controller is connected to the output terminal of the local edge control node to receive the photovoltaic power prediction curve and the grid dispatch command. The centralized coordination controller establishes a rolling optimization objective function based on the model predictive control algorithm, which includes energy storage charging and discharging power constraints and upper and lower limit constraints of state of charge. By solving the rolling optimization objective function, the energy storage charging and discharging power reference value for the next control cycle is output. The communication terminal of the energy storage converter is connected to the output terminal of the centralized coordination controller to receive the energy storage charging and discharging power reference value. The energy storage converter adjusts the DC bus voltage on the energy storage side according to the energy storage charging and discharging power reference value using a composite control strategy combining feedforward compensation and proportional-integral feedback to perform charging and discharging actions.

2. The distributed photovoltaic energy storage power generation system according to claim 1, characterized in that, The local edge control node includes a meteorological data preprocessing sub-node and a prediction model inference sub-node; The meteorological data preprocessing sub-node performs time alignment and missing value imputation on the environmental meteorological data, and divides the processed environmental meteorological data into meteorological feature sequences according to a preset time window. The prediction model inference sub-node embeds a long short-term memory network. The prediction model inference sub-node inputs the meteorological feature sequence into the long short-term memory network, calculates the cell state update value and the hidden layer state vector through the forget gate, input gate and output gate of the long short-term memory network, and maps the hidden layer state vector to the photovoltaic power prediction curve. The local edge control node packages the photovoltaic power prediction curve into a timestamped data frame and sends it to the centralized coordination controller.

3. A distributed photovoltaic energy storage power generation system according to claim 1, characterized in that, The centralized coordination controller includes an optimization model construction sub-node and a solver sub-node; The optimization model construction sub-node establishes an objective function term with the optimization objectives of minimizing the power fluctuation at the grid connection point and minimizing the frequency of energy storage charging and discharging actions. The optimization model construction sub-node establishes inequality constraint terms with the upper limit of energy storage charging and discharging power, the lower limit of energy storage charging and discharging power, the current state of charge value, the upper limit of state of charge, and the lower limit of state of charge as constraints. The optimization model construction sub-node combines the objective function term and the inequality constraint terms to generate a rolling optimization objective function. The solver sub-node calls the interior point method to iteratively optimize the rolling optimization objective function and outputs the energy storage charging and discharging power sequence in the future preset control time domain. The solver sub-node extracts the first value in the energy storage charging and discharging power sequence as the energy storage charging and discharging power reference value for the next control cycle.

4. A distributed photovoltaic energy storage power generation system according to claim 1, characterized in that, The energy storage converter includes a reference value analysis sub-node, a feedforward control loop sub-node, and a proportional-integral feedback control loop sub-node. The reference value parsing sub-node performs a table lookup and conversion on the energy storage charging and discharging power reference value to obtain the reference value of DC current on the energy storage side. The feedforward control loop sub-node collects the DC bus voltage and DC current of the energy storage side, calculates the current instantaneous power based on the DC bus voltage and DC current of the energy storage side, and inputs the difference between the current instantaneous power and the energy storage charging and discharging power reference value as the feedforward disturbance compensation quantity to the feedforward control loop sub-node. The proportional-integral feedback control loop sub-node uses the difference between the reference value of the DC current on the energy storage side and the DC current on the energy storage side as the input of the proportional-integral regulator. The output of the proportional-integral regulator is superimposed with the feedforward disturbance compensation to generate a pulse width modulation signal to drive the power switching tube of the energy storage converter.

5. A distributed photovoltaic energy storage power generation system according to claim 1, characterized in that, The local edge control node establishes a communication link based on the User Datagram Protocol (UDP) with the centralized coordination controller. The local edge control node adds a frame type identifier and a sequence number to the frame header of the data frame and adds a cyclic redundancy check (CRC) code to the frame tail of the data frame. After receiving a data frame, the centralized coordination controller parses the frame type identifier. When the frame type identifier indicates a photovoltaic power prediction curve, the centralized coordination controller extracts the sequence number and performs time-series rearrangement to restore the time order of the photovoltaic power prediction curve. A communication link based on a controller local area network is established between the centralized coordination controller and the energy storage converter. The centralized coordination controller sends controller local area network messages containing the energy storage charging and discharging power reference value to the energy storage converter at preset equal time intervals.

6. A distributed photovoltaic energy storage power generation system according to claim 1, characterized in that, The centralized coordination controller has a built-in communication watchdog timer and a degradation control trigger sub-node; When the communication watchdog timer detects that the local edge control node has continuously lost a preset number of photovoltaic power prediction curve data frames, the communication watchdog timer outputs an abnormal trigger signal to the degraded control trigger sub-node. Upon receiving an abnormal trigger signal, the degradation control trigger sub-node masks the prediction correlation term in the rolling optimization objective function and switches the optimization objective function to a constant power control mode that tracks the grid dispatch command. The degradation control trigger sub-node calculates the degradation energy storage charging and discharging power reference value based on the current energy storage state of charge value and the grid dispatch command, and sends the degradation energy storage charging and discharging power reference value to the energy storage converter.

7. A distributed photovoltaic energy storage power generation system according to claim 2, characterized in that, The prediction model inference sub-node is connected to the model online update sub-node, which stores historical meteorological feature sequences and historical actual photovoltaic output power. The model updates its sub-nodes online according to a preset training period, inputs historical meteorological feature sequences into a long short-term memory network to obtain historical predicted power, and calculates the mean square error loss value between historical predicted power and historical actual photovoltaic output power. The model online update sub-node uses the gradient descent algorithm to calculate the gradient of the weight matrix of each network layer in the long short-term memory network based on the mean squared error loss value, and uses the backpropagation mechanism to update the weight matrix. The model online update sub-node synchronizes the updated weight matrix to the prediction model inference sub-node.

8. A distributed photovoltaic energy storage power generation system according to claim 3, characterized in that, The optimization model construction sub-node is also connected to the charge state dynamic boundary calculation sub-node; The state of charge dynamic boundary calculation sub-node collects ambient temperature data and cycle count data of the energy storage battery in real time. Based on the ambient temperature data, it queries the preset temperature-internal resistance mapping table to obtain the current internal resistance value of the battery. The state of charge dynamic boundary calculation sub-node combines the current internal resistance value of the battery with the cycle count data to calculate the maximum usable capacity decay coefficient of the energy storage battery. The optimization model constructs sub-nodes that downwardly adjust the upper limit and lower limit of the initial state of charge based on the maximum available capacity decay coefficient, and use the adjusted upper limit and lower limit of the state of charge as the upper and lower limit constraints of the state of charge in the rolling optimization objective function.

9. A distributed photovoltaic energy storage power generation system according to claim 4, characterized in that, The proportional-integral feedback control loop sub-nodes are embedded with anti-integral saturation processing logic. When the output of the proportional-integral regulator reaches the preset output limit value, the anti-integral saturation processing logic freezes the integral accumulation operation inside the proportional-integral regulator and calculates the product of the difference between the reference value of the DC current on the energy storage side and the DC current on the energy storage side and the proportional gain. When the product is less than the preset output limit and the polarity of the difference is reversed, the anti-integral saturation processing logic is unfrozen and restarts the integral accumulation operation according to the current difference. The proportional-integral feedback control loop sub-node generates the output of the proportional-integral regulator based on the integral accumulation result after the restart.

10. A distributed photovoltaic energy storage power generation control method, applied to the distributed photovoltaic energy storage power generation system described in claim 1, characterized in that, The method includes: Step 1, collecting instantaneous photovoltaic output power and environmental meteorological data of the photovoltaic power generation array through a local edge control node, and using a long short-term memory network to extract features from the environmental meteorological data to generate a photovoltaic power prediction curve within a preset time period; Step 2: Receive photovoltaic power prediction curves and grid dispatch instructions through a centralized coordination controller. Establish a rolling optimization objective function based on the model predictive control algorithm, which includes energy storage charging and discharging power constraints and upper and lower limits of state of charge constraints. Solve the rolling optimization objective function to output the energy storage charging and discharging power benchmark value for the next control cycle. Step 3: Receive the energy storage charging and discharging power reference value through the energy storage converter, and adjust the DC bus voltage on the energy storage side according to the energy storage charging and discharging power reference value using a composite control strategy combining feedforward compensation and proportional-integral feedback to perform charging and discharging actions.

Images