A quantum-accelerated maximum power point tracking method for partially shaded photovoltaic multi-peaks

Abstract

This paper proposes a quantum-accelerated maximum power point tracking (MPPT) method for partially shaded photovoltaic multi-peaks. This method uses the MobileNet V3 network in photovoltaic power generation control technology to predict the maximum power point of photovoltaic modules under current shading conditions. The Grover algorithm is then used to adjust the parameters of a fractional-order active disturbance rejection controller (ADRC) to achieve MPPT tracking, ultimately improving photovoltaic power generation efficiency. This method addresses the slow tracking speed and severe steady-state oscillations inherent in existing photovoltaic MPPT technologies. It implements MPPT technology, optimizes the maximum power point search process, and controls the operating point tracking of the MPPT, ultimately improving photovoltaic power generation efficiency.

Description

Technical Field

[0001] The present invention belongs to the fields of power systems, microgrids, smart grids, new power systems, and new energy, and involves deep learning, quantum acceleration search, and control methods. It is suitable for tracking and controlling the maximum power point of multi-peak scenarios of photovoltaic power generation under partial shading conditions. Background Art

[0002] Existing maximum power point tracking technologies (proportional-integral-differential, auto-disturbance rejection control, fuzzy control, artificial neural networks, model predictive control, deep neural networks, and deep reinforcement learning) suffer from slow tracking speeds and severe steady-state oscillations. Traditional control methods (proportional-integral-differential, auto-disturbance rejection control, and fuzzy control) offer precise and fast control, but their parameters require optimization algorithms. Deep learning methods (artificial neural networks, model predictive control, and deep neural networks) do not offer the same control accuracy as traditional methods. Deep reinforcement learning methods (such as Q-learning, SARSA, PPO, and TRPO) require offline training or suffer from the high trial-and-error nature of online processes, resulting in low accuracy.

[0003] Furthermore, existing maximum power point tracking (MPPT) technologies are prone to mistracking in rapidly changing environments. Optimization algorithms, such as cuckoo optimization and particle swarm optimization, are typically used to optimize the parameters of these tracking control methods. These algorithms require a lengthy iterative process.

[0004] Therefore, a quantum accelerated maximum power point tracking method with partially shielded photovoltaic multi-peaks is proposed to solve the problems of slow tracking speed, severe steady-state oscillation and mistracking in the maximum power point tracking technology. Summary of the Invention

[0005] This paper proposes a quantum-accelerated maximum power point tracking method for partially obscured photovoltaic multi-peaks. It combines the MobileNet V3 network, the Grover algorithm, and fractional-order active disturbance rejection control for tracking the maximum power point in photovoltaic power generation technology. It can improve the speed and dynamic performance of photovoltaic system maximum power point tracking. The steps in the use process are as follows:

[0006] Step (1): measuring the irradiance and temperature under current environmental conditions;

[0007] Use irradiance sensors and temperature sensors to collect irradiance and temperature under current environmental conditions. Since the computer stores image data in the form of a 3D array in the digital image processing program, the historical data of irradiance and temperature obtained are formed into a matrix. The irradiance and temperature matrix is:

[0008]

[0009] Where E is the irradiance data matrix; ET is the irradiance sensor; ET(t) is the irradiance collected by the irradiance sensor in real time; T is the temperature data matrix; TT is the temperature sensor; TT(t) is the temperature collected by the temperature sensor in real time; D is the matrix formed by the irradiance and temperature historical data; image() is a function that forms a 3D array image matrix from historical data, and the matrix size is 224×224×3;

[0010] Step (2): Use the MobileNet V3 network in the convolutional neural network to estimate the maximum power point reference value of the photovoltaic module;

[0011] The input of the MobileNet V3 network is the image formed by the historical irradiance and temperature obtained in step (1). The output of the MobileNet V3 network is the reference voltage and reference current of the maximum power point of the photovoltaic module predicted by the network. The obtained maximum power point reference value of the photovoltaic module will be used to further estimate the maximum power point reference value of the entire photovoltaic array. The MobileNet V3 network prediction model is:

[0012] MPP(V MPP ,I MPP )=MobileNetV3(D) (2)

[0013] Where MobileNetV3 is the MobileNet V3 network in the convolutional neural network; MPP is the maximum power point; V MPP and I MPP The reference voltage and reference current of the maximum power point of the photovoltaic module predicted by the network;

[0014] Step (3): Combine the Boost converter discrete model with the general photovoltaic array discrete model to establish a photovoltaic system discrete prediction model;

[0015] The maximum power point reference value of the photovoltaic array is estimated by combining the maximum power point reference value of the photovoltaic module predicted by the MobileNet V3 network;

[0016] The Boost circuit model under the condition that the switching element is turned on is:

[0017]

[0018] Where, I L is the inductor current; is the time differential of the inductor current; V array is the PV array voltage; is the differential of the PV array voltage with respect to time; I array is the photovoltaic array current; L is the Boost circuit inductance; C is the Boost circuit capacitance;

[0019] The discrete mathematical model after discretization is:

[0020]

[0021] Where k is the ordinal number in the discrete expression; S is the state of the switch element, S = 1 means the switch is in the on state; S = 0 means the switch is in the off state; T S is the discrete sampling period; I L (k) is the inductor current at time k; I L (k+1) is the inductor current at time k+1; V array (k) is the PV array voltage at time k; V array (k+1) is the PV array voltage at time k+1; I array (k) is the PV array current at time k;

[0022] The Boost circuit model under the cut-off condition of the switching element is:

[0023]

[0024] Where V DC is the equivalent DC source voltage;

[0025] The discretized mathematical model after discretization is:

[0026]

[0027] The discrete model of the Boost converter is obtained from equations (2) and (3):

[0028]

[0029] A discrete prediction model for photovoltaic modules based on a dynamic conductivity model is used. The dynamic conductivity model is:

[0030] I PV (k+1)=I PV (k)+G PV (k)k·(V PV (k+1)-V PV (k)) (8)

[0031] Where V PV (k) and I PV (k) is the voltage and current of the photovoltaic module at time k; V PV (k+1) and I PV (k+1) is the voltage and current of the photovoltaic module at time k+1; G PV (k) is the dynamic conductivity of the photovoltaic module at time k;

[0032] The model of dynamic conductivity is:

[0033]

[0034] Where V PV (k-1) and I PV (k-1) is the voltage and current of the PV module at time k-1;

[0035] The discrete model of a general photovoltaic array is:

[0036] I array (k+1)=G array (k)(V array (k+1)-V array (k))+I array (k) (10)

[0037] Where G array (k) is the dynamic conductivity of the photovoltaic array at time k;

[0038] Combining the discrete model of the Boost converter with the discrete model of the general photovoltaic array, the discrete prediction model of the photovoltaic system is obtained as follows:

[0039]

[0040] Finally, combined with the maximum power point reference value of the photovoltaic module obtained in step (2), the maximum power point reference voltage V of the photovoltaic array is obtained. MPP and reference current I MPP ;

[0041] Step (4): Use the quantum accelerated Grover method to tune the parameters of the fractional-order active disturbance rejection controller;

[0042] The fractional-order active disturbance rejection controller consists of three parts: a fractional-order proportional-differential (FOPD) controller, a linear extended state observer (LESO), and a linear feedback control law consisting of a disturbance compensation b0.

[0043] The transfer function of the fractional-order proportional-derivative FOPD controller is:

[0044] C(s)=K p +K d s μ ,0<μ<2 (12)

[0045] Where K p is the proportionality coefficient; K d is the differential coefficient; s μ is a fractional differential;

[0046] The state equation of the linear extended state observer LESO is:

[0047]

[0048] Where ω0 is the observer bandwidth in the bandwidth method; ω c is the controller bandwidth in the bandwidth method; u is the input of the linear extended state observer LESO; y is the output of the linear extended state observer LESO; x is the observation quantity of the linear extended state observer LESO;

[0049] The disturbance compensation link is:

[0050]

[0051] Where v is the given quantity of the system, i.e., the expected input value; z1, z2, and z3 are the states observed by the linear extended state observer LESO; z1 = y is the tracking signal of y. is the differential signal of y, z3 is the state quantity expanded by LESO to the controlled object; b0 is the compensation factor;

[0052] Grover acceleration method is:

[0053]

[0054] Where O is the Oracle operator; G is the Grover operator; is a quantum state representation; is the right vector of the Dirac symbol for the quantum state, and is the initial state of each quantum bit; is the right vector of the Dirac symbol representation of the quantum state; I is the identity matrix;

[0055] The control parameters of the fractional-order ADRC to be tuned are K p , K d , μ, ω0, ω c and b0; the independent distributed parameter tuning method based on the bandwidth method and the Grover acceleration method is used to complete the tuning of the six parameters of the fractional-order active disturbance rejection controller; the parameters of the fractional-order active disturbance rejection controller are divided into K p , K d The fractional order parameters of μ and μ include ω0, ω c The linear auto-disturbance rejection parameters of and b0 are adjusted independently.

[0056] First, the bandwidth method is used to adjust the parameters of the linear ADRC part, and ω0, ω c and b0 parameter value;

[0057] According to the stability conditions of the fractional-order controller, that is, satisfying equations (12), (13) and (14), the Grover acceleration method is used to calculate the fractional-order parameter K. p , K d and μ are tuned; the Grover acceleration method searches for quantum states, marking the target item for phase flipping and amplifying the probability amplitude, and measurement steps; if the current search successfully finds the target item, i.e., the optimal parameter, the search ends; otherwise, the search continues until the target item is found; after the search is completed, the optimal fractional-order parameter K is obtained. p , K d and μ, which completes the parameter tuning of the fractional-order active disturbance rejection controller;

[0058] Step (5): Set the maximum power point reference voltage V of the photovoltaic array obtained in step (3) to MPP and reference current I MPP The given quantity, i.e., the expected input value, is input into the fractional-order active disturbance rejection controller tuned by the Grover acceleration method in step (4), and is applied to the optimization process of finding the maximum power point in the photovoltaic power generation system and the control process of controlling the operating point to track the maximum power point.

[0059] The present invention has the following advantages and effects compared to the prior art:

[0060] (1) The MobileNet V3 convolutional neural network processes ambient temperature and irradiance to obtain the maximum power point reference value of the photovoltaic module, which is then used to further estimate the maximum power point reference value of the entire photovoltaic array. The MobileNet V3 network is data-driven; the more data, the higher the prediction accuracy of the MobileNet V3 network. Compared with existing prediction models, deep neural networks are faster, have no oscillation problems, and have higher prediction accuracy.

[0061] (2) Most existing maximum power point tracking technology control methods use PID controllers in their control process. Although PID controllers do not rely on models, they have poor dynamic performance and are difficult to meet the requirements of fast tracking of the maximum power point under rapidly changing environmental conditions. The present invention adopts a fractional-order active disturbance rejection controller, which retains the advantages of the PID controller while improving its disadvantages. It has a faster response, higher control accuracy, stronger active disturbance rejection robustness, and stronger anti-interference ability. Compared with the PID controller, the method proposed in the present invention takes into account fractional-order and more dimensional information. Compared with the fractional-order PID controller, the method proposed in the present invention takes into account more dimensional information. Compared with the active disturbance rejection controller, the method proposed in the present invention takes into account fractional-order information.

[0062] (3) The performance of the fractional-order ADRC controller is heavily dependent on parameter tuning and optimization. The Grover algorithm in quantum acceleration search is used to tune the parameters of the fractional-order ADRC controller. However, the addition of differential order parameters makes the controller parameter tuning more complex. The Grover algorithm is used to tune the parameters of the fractional-order part of the fractional-order ADRC controller, reducing the tuning time and enabling accurate search for the optimal parameters. Compared to deep learning methods and deep reinforcement learning methods, the method proposed in this paper has an acceleration effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a framework diagram of a maximum power point tracking system for a photovoltaic power generation system according to the method of the present invention.

[0064] Figure 2 This is a flow chart of the MobileNet V3 network structure of the method of the present invention.

[0065] Figure 3 It is a structural flow chart of the Grover acceleration method of the method of the present invention.

[0066] Figure 4 It is a quantum circuit diagram prepared by the Grover operator of the method of the present invention.

[0067] Figure 5 This is a structural flow chart of the fractional-order active disturbance rejection controller of the method of the present invention. DETAILED DESCRIPTION

[0068] The present invention proposes a quantum accelerated maximum power point tracking method for partially shielding photovoltaic multi-peaks, which is described in detail with reference to the accompanying drawings as follows:

[0069] Figure 1 This is a framework diagram of the maximum power point tracking system for photovoltaic power generation systems according to the present invention. First, the historical data of irradiance E and temperature T collected by the irradiance sensor and temperature sensor are formed into a digital image D. The digital image D is input into the trained MobileNet V3 network, which outputs the maximum power point reference value of the photovoltaic module. Combined with the photovoltaic array prediction model, the maximum power point voltage reference value V is obtained. MPP and current reference value I MPP ; Then, the predicted maximum power point voltage reference value V MPP and current reference value I MPPThe deviation is compared with the actual maximum power point voltage value V and the current reference value I, and the deviation and the state quantities z1, z2 and z3 observed by the linear extended observer are input into the fractional-order active disturbance rejection controller, where z1 is the tracking signal of y, z2 is the differential signal of y, and z3 is the state quantity expanded by LESO to the controlled object; then, the fractional-order part parameter K of the fractional-order active disturbance rejection controller is adjusted using the Grover algorithm. p , K d and μ, and the remaining linear ADRC parameters ω0, ω c and b0 are adjusted using the conventional bandwidth method; finally, the fractional-order active disturbance rejection controller controls the photovoltaic panel to adjust and output the actual power point.

[0070] Figure 2 This is a flowchart of the MobileNet V3 network structure of the method of the present invention. "conv2d" represents a normal convolution, and "bneck" represents a linear bottleneck. First, an irradiance sensor ET and a temperature sensor TT are used to collect the irradiance E and temperature T of the external environment. Historical data of irradiance E and temperature T are used to form a digital image D of size 224×224×3. This digital image D is then input into the input layer of a trained MobileNet V3 network. Next, digital image D passes through the network in sequence: a convolutional layer, a Reluctant Inverter (ReLU) activation layer, an average pooling layer, a perception model, a convolutional layer, a Reluctant Inverter (ReLU) activation layer, a pooling layer, a dropout layer, a fully connected layer, a fully connected layer, and a sigmoid activation layer. This involves sampling digital image D five times, so that feature layers of different sizes contain feature information at different levels. Finally, this top-down network path fuses low-level, detailed features of the image semantic information formed by irradiance E and temperature T with high-level, abstract features, thereby predicting the reference voltage and reference current at the maximum power point of the photovoltaic module.

[0071] Figure 3 The Grover acceleration method is used to calculate the fractional order parameter K. p , K d and μ are tuned; the Grover acceleration method searches for quantum states, marking the target item for phase flipping and amplifying the probability amplitude, and measurement steps; if the current search successfully finds the target item, i.e., the optimal parameter, the search ends; otherwise, the search continues until the target item is found; after the search is completed, the optimal fractional-order parameter K is obtained. p , K d and μ, thus completing the parameter tuning of the fractional-order active disturbance rejection controller.

[0072] Figure 4This is the quantum circuit diagram prepared by the Grover operator of the method of the present invention. The function of the Grover operator is to amplify the probability of the target state of interest. The Grover operator consists of two parts. The Oracle operator is used to mark the quantum state of interest, and then the diffusion operator is used to amplify the probability of the marked state. If n quantum bits are input, the database size at this time is N = 2 n First, a Hadamard gate H is added to each bit of the initial state to obtain a uniform superposition state; then, the phase is flipped to mark the target item, making the phase of the target item negative; finally, the mean is flipped to move the initial state closer to the position of the target item; through multiple Grover operators connected in series, multiple quantum states can be amplified to varying degrees to achieve the desired effect.

[0073] Figure 5 The flow chart of the fractional order ADRC structure of the method of the present invention is as follows: First, the input of the fractional order ADRC is the predicted maximum power point reference value MPP. * Subtract the difference obtained from the measured controlled variable MPP and the measured controlled variable MPP. At the beginning, the larger error control signal e(t) can activate the object quickly; then, MPP * The difference e(t) obtained by the -MPP is passed through the FOPD controller and then subtracted from the total disturbance value z2 output by the fractional-order linear extended state observer (LESO) to obtain the difference e1(t). The state quantity z3 expanded by the fractional-order linear extended state observer (LESO) for the controlled object is subtracted from e1(t). The difference e2(t) obtained is passed through 1 / b0 to obtain the output control law. Finally, the controlled object is controlled according to the output control law, and the fractional-order linear extended state observer (LESO) observes the actual output maximum power point (MPP) in real time.

[0074] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A quantum accelerated maximum power point tracking method for partially shielding photovoltaic multi-peaks, characterized in that: Combining the MobileNetV3 network, Grover algorithm, and fractional-order active disturbance rejection control for maximum power point tracking in photovoltaic power generation technology can improve the speed and dynamic performance of photovoltaic system tracking maximum power point. The steps in the use process are: Step (1): measuring the irradiance and temperature under current environmental conditions; Use irradiance sensors and temperature sensors to collect irradiance and temperature under current environmental conditions. Since the computer stores image data in the form of a 3D array in the digital image processing program, the historical data of irradiance and temperature obtained are formed into a matrix. The irradiance and temperature matrix is: Where E is the irradiance data matrix; ET is the irradiance sensor; ET(t) is the irradiance collected by the irradiance sensor in real time; T is the temperature data matrix; TT is the temperature sensor; TT(t) is the temperature collected by the temperature sensor in real time; D is the matrix formed by the irradiance and temperature historical data; image() is a function that forms a 3D array image matrix from historical data, and the matrix size is 224×224×3; Step (2): Use the MobileNet V3 network in the convolutional neural network to estimate the maximum power point reference value of the photovoltaic module; The input of the MobileNet V3 network is the image formed by the historical irradiance and temperature obtained in step (1). The output of the MobileNet V3 network is the reference voltage and reference current of the maximum power point of the photovoltaic module predicted by the network. The obtained maximum power point reference value of the photovoltaic module will be used to further estimate the maximum power point reference value of the entire photovoltaic array. The MobileNet V3 network prediction model is: MPP(V MPP ,I MPP )=MobileNetV3(D) (2) Where MobileNetV3 is the MobileNet V3 network in the convolutional neural network; MPP is the maximum power point; V MPP and I MPP The reference voltage and reference current of the maximum power point of the photovoltaic module predicted by the network; Step (3): Combine the Boost converter discrete model with the general photovoltaic array discrete model to establish a photovoltaic system discrete prediction model; The maximum power point reference value of the photovoltaic array is estimated by combining the maximum power point reference value of the photovoltaic module predicted by the MobileNet V3 network; The Boost circuit model under the condition that the switching element is turned on is: Where, I L is the inductor current; is the time differential of the inductor current; V array is the PV array voltage; is the differential of the PV array voltage with respect to time; I array is the photovoltaic array current; L is the Boost circuit inductance; C is the Boost circuit capacitance; The discrete mathematical model after discretization is: Where k is the ordinal number in the discrete expression; S is the state of the switch element, S = 1 means the switch is in the on state; S = 0 means the switch is in the off state; T S is the discrete sampling period; I L (k) is the inductor current at time k; I L (k+1) is the inductor current at time k+1; V array (k) is the PV array voltage at time k; V array (k+1) is the PV array voltage at time k+1; I array (k) is the PV array current at time k; The Boost circuit model under the cut-off condition of the switching element is: Where V DC is the equivalent DC source voltage; The discretized mathematical model after discretization is: The discrete model of the Boost converter is obtained from equations (2) and (3): A discrete prediction model for photovoltaic modules based on a dynamic conductivity model is used. The dynamic conductivity model is: I PV (k+1)=I PV (k)+G PV (k)·(V PV (k+1)-V PV (k)) (8) Where V PV (k) and I PV (k) is the voltage and current of the photovoltaic module at time k; V PV (k+1) and I PV (k+1) is the voltage and current of the photovoltaic module at time k+1; G PV (k) is the dynamic conductivity of the photovoltaic module at time k; The model of dynamic conductivity is: Where V PV (k-1) and I PV (k-1) is the voltage and current of the PV module at time k-1; The discrete model of a general photovoltaic array is: I array (k+1)=G array (k)(V array (k+1)-V array (k))+I array (k) (10) In the formula, G array (k) is the dynamic conductivity of the photovoltaic array at time k; Combining the discrete model of the Boost converter with the discrete model of the general photovoltaic array, the discrete prediction model of the photovoltaic system is obtained as follows: Finally, combined with the maximum power point reference value of the photovoltaic module obtained in step (2), the maximum power point reference voltage V of the photovoltaic array is obtained. MPP and reference current I MPP ; Step (4): Use the quantum accelerated Grover method to tune the parameters of the fractional-order active disturbance rejection controller; The fractional-order active disturbance rejection controller consists of three parts: a fractional-order proportional-differential (FOPD) controller, a linear extended state observer (LESO), and a linear feedback control law consisting of a disturbance compensation b0. The transfer function of the fractional-order proportional-derivative FOPD controller is: C(s)=K p +K d s μ ,0<μ<2 (12) Where K p is the proportionality coefficient; K d is the differential coefficient; s μ is a fractional differential; The state equation of the linear extended state observer LESO is: Where ω0 is the observer bandwidth in the bandwidth method; ω c is the controller bandwidth in the bandwidth method; u is the input of the linear extended state observer LESO; y is the output of the linear extended state observer LESO; x is the observation of the linear extended state observer LESO; The disturbance compensation link is: Where v is the given quantity of the system, i.e., the expected input value; z1, z2, and z3 are the states observed by the linear extended state observer LESO; z1 = y is the tracking signal of y. is the differential signal of y, z3 is the state quantity expanded by LESO to the controlled object; b0 is the compensation factor; Grover acceleration method is: Where O is the Oracle operator; G is the Grover operator; is a quantum state representation; is the right vector of the Dirac symbol for the quantum state, and is the initial state of each quantum bit; is the right vector of the Dirac symbol representation of the quantum state; I is the identity matrix; The control parameters of the fractional-order ADRC to be tuned are K p , K d , μ, ω0, ω c and b0; the independent distributed parameter tuning method based on the bandwidth method and the Grover acceleration method is used to complete the tuning of the six parameters of the fractional-order active disturbance rejection controller; the parameters of the fractional-order active disturbance rejection controller are divided into K p , K d The fractional order parameters of μ and μ include ω0, ω c The linear auto-disturbance rejection parameters of and b0 are adjusted independently. First, the bandwidth method is used to adjust the parameters of the linear ADRC part, and ω0, ω c and b0 parameter value; According to the stability conditions of the fractional-order controller, that is, satisfying equations (12), (13) and (14), the Grover acceleration method is used to calculate the fractional-order parameter K. p , K d and μ are tuned; the Grover acceleration method searches for quantum states, marking the target item for phase flipping and amplifying the probability amplitude, and measurement steps; if the current search successfully finds the target item, i.e., the optimal parameter, the search ends; otherwise, the search continues until the target item is found; after the search is completed, the optimal fractional-order parameter K is obtained. p , K d and μ, which completes the parameter tuning of the fractional-order active disturbance rejection controller; Step (5): Set the maximum power point reference voltage V of the photovoltaic array obtained in step (3) to MPP and reference current I MPP The given quantity, i.e., the expected input value, is input into the fractional-order active disturbance rejection controller tuned by the Grover acceleration method in step (4), and is applied to the optimization process of finding the maximum power point in the photovoltaic power generation system and the control process of controlling the operating point to track the maximum power point.

Images