An environmentally adaptive photovoltaic current observation method based on neural network

Abstract

The present invention relates to an environmentally adaptive photovoltaic current observation method based on a neural network, and belongs to the field of control technology in power electronics technology. In the prior art, there are problems such as the observation model relying on circuit parasitic parameters or load equivalent resistance, the photovoltaic system lacking constant current or constant power output capability, or still requiring additional environmental sensors such as illumination and temperature. The method described in the present invention realizes online real-time and accurate observation of photovoltaic current by accurately identifying the equivalent loss resistance of the Boost DC converter, thereby avoiding the problems of large loss, susceptibility to interference and high cost of current sensors in the photovoltaic storage DC microgrid system. There is no need to obtain accurate circuit component parameters or environmental sensors, only voltage sampling circuits on the input and output sides of the converter are required, and offline training of the neural network is performed using sample data, so as to realize environmentally adaptive photovoltaic current observation. Thereby significantly reducing the volume and economic cost of the photovoltaic system, and further improving the comprehensive performance of the photovoltaic system.

Description

Technical Field

[0001] The present invention relates to the field of control technology in power electronics technology, and in particular to an environment-adaptive photovoltaic current observation method based on a neural network. Background Art

[0002] In the photovoltaic-storage DC microgrid system, the distributed photovoltaic arrays are connected to the DC bus through the Boost DC converter to supply energy to the energy storage battery and the load. Therefore, the stability and energy utilization efficiency of the photovoltaic-storage DC microgrid system largely depend on the efficient operation and control of the Boost DC converter. Many literatures have conducted in-depth research on the photovoltaic maximum power point tracking technology (MPPT), and achieved photovoltaic operation at the maximum power operating point through constant voltage method, perturbation observation method, incremental conductance method and other optimization intelligent algorithms. In addition, in response to the operation requirements of the photovoltaic-storage DC microgrid system, some literatures have proposed a multi-mode photovoltaic-storage DC microgrid system operation control method to maintain a stable system DC bus voltage and achieve precise control of photovoltaic output power.

[0003] The above control strategy needs to sample the photovoltaic voltage and current and the DC bus voltage, and then generate a control signal for the power switch of the Boost DC converter after inputting it into the digital controller through the analog-to-digital conversion circuit (ADC), so as to realize the control of the energy flow of the photovoltaic storage DC microgrid system. Among them, the voltage sampling circuit is usually composed of a parallel voltage divider resistor and an isolation operational amplifier, while the current sampling needs to be realized by a Hall sensor or a series sampling resistor. The Hall sensor isolates and samples the current through the magnetic field generated by the current. However, on the one hand, the Hall sensor is expensive and is prone to demagnetization when the current is too large. On the other hand, it is very sensitive to external magnetic fields and is easily affected by external or stray magnetic fields. The high-frequency switching action of the power switch in the Boost DC converter will superimpose noise on the current sampling signal, causing current sampling errors. Using a series resistor for sampling increases the loss of the power circuit and generates additional heat and voltage drop. If a series sampling resistor with a smaller resistance is used, although the loss is relatively reduced, its terminal voltage is also smaller, which puts higher requirements on the performance of the subsequent operational amplifier. In addition, since the sampling resistor is connected in series with the power circuit, once the sampling resistor fails and opens, the photovoltaic array will be directly disconnected from the DC bus. In summary, the existing current sampling schemes have limitations in terms of cost, efficiency, reliability, etc. Therefore, studying the current observation method without current sampling circuit is of great significance to improve the comprehensive performance of the photovoltaic energy storage DC microgrid system.

[0004] References to related patent applications:

[0005] Existing current observation technology establishes a mathematical model of the inductor current and capacitor voltage in the circuit according to the converter operation mode, and only uses a voltage sensor to achieve current observation. [1] invented a power converter current detection method based on non-smooth observation technology. By sampling the output voltage of the circuit and establishing a mathematical model of the main circuit, the current observer is constructed using non-smooth technology to achieve the detection of the inductor current in the Buck converter. [2] invented a DAB converter current sensorless control method based on model prediction. By obtaining the input and output voltages of the converter and combining the phase shift angle output by the current compensator, the model prediction of the load current is achieved. [3] invented a single-sensor photovoltaic module optimizer and its control method. By only detecting the converter output voltage and then adjusting the duty cycle, the photovoltaic maximum power operating point is tracked.

[0006] The existing current sensorless observation methods have the following three main problems: (1) The observation model relies on circuit parasitic parameters or load equivalent resistance to participate in the calculation, but in practical applications, parasitic parameters are often difficult to obtain accurately, and load equivalent resistance is volatile and unpredictable. (2) The current observer can meet the voltage control or MPPT control performance, but there is an error between the current observation value and the actual value, which makes the photovoltaic system lack constant current or constant power output capability. (3) Although the current sampling circuit can be eliminated, additional environmental sensors such as light and temperature are required to participate in the control.

[0007] [1] Ma Li, Ni Yuanyuan, Ding Shihong. Current sensorless control method for DAB converter based on model prediction[P]. Shaanxi Province: CN106526299B, 2019-03-05.

[0008] [2] Wu Xuanluo, Tu Wencong, Jiang Yu, Yang Weiping. Non-isolated low current ripple high voltage gain soft switching DC-DC converter [P]. Tianjin: CN114844364A, 2022-08-02.

[0009] [3] Bi Qiang, Zhou Guohua, Leng Minrui, Tian Qingxin. A single sensor photovoltaic module optimizer and its control method [P]. Sichuan Province: CN109546961B, 2020-09-01. Summary of the invention

[0010] In view of the defects existing in the prior art, the purpose of the present invention is to achieve real-time and accurate observation of photovoltaic current, avoid the problems of large loss, susceptibility to interference and high cost of current sensors, and further improve the comprehensive performance of photovoltaic systems.

[0011] To achieve the above purpose, the technical solution adopted by the present invention is: an environment-adaptive photovoltaic current observation method based on a neural network, comprising the following steps:

[0012] (1) Set a load resistor R of known resistance on the output side of the Boost DC converter load , calculate the equivalent loss resistance value R under the corresponding working conditions of each environment EQ , together with the voltage sampling data and duty cycle, constitute a sample set as the training data of the neural network;

[0013] (2) Construct and train a neural network. When the photovoltaic system is running, use the obtained neural network model to perform online identification of the converter equivalent loss resistance under various working conditions to obtain the identification value R of the equivalent loss resistance. EQ *;

[0014] (3) Using the identified equivalent loss resistance R EQ *Carry out online real-time observation of photovoltaic current to achieve efficient operation control of photovoltaic system.

[0015] Furthermore, in step (1), the sample data set is obtained based on actual circuit and voltage sampling. During the data acquisition process, the duty cycle of the power switch is changed, and the photovoltaic voltage U under each working condition is pv , bus voltage U bus Sampling values ​​for data collection.

[0016] Furthermore, in step (1), the sample data set of the neural network consists of the photovoltaic voltage U pv , bus voltage U bus , duty cycle D, slope of the IV output characteristic curve of the photovoltaic cell And the equivalent loss resistance value R of the converter EQ constitute.

[0017] Furthermore, the equivalent loss resistance R in the sample data set EQ The power switch duty cycle D and photovoltaic voltage U pv , bus voltage U bus and bus load value R load Calculation, R EQ The mathematical model is shown as follows:

[0018]

[0019] Wherein, D' is the complementary value of the duty cycle, that is, D'=1-D.

[0020] Furthermore, the slope of the IV output characteristic curve of the photovoltaic cell in the sample data set is The photovoltaic voltage U pv , bus voltage U bus , the duty cycle D is calculated according to the following formula:

[0021]

[0022] Wherein, k is the number of sampling cycles, and D' is the complementary value of the duty cycle, that is, D'=1-D.

[0023] Furthermore, in step (2), a power switch duty cycle D, photovoltaic voltage U pv , bus voltage U bus , the slope of the IV output characteristic curve of the photovoltaic cell is the input, the equivalent loss resistance R EQ A neural network structure with four inputs and one output.

[0024] Furthermore, the specific internal structure of the neural network is determined by the actual training effect. Based on the sample data set obtained from the experiment, neural networks with different structures are trained multiple times, and the training results are evaluated and compared using the error function. The neural network with the best result is selected as the equivalent loss resistor R EQ The identification model is embedded into the digital controller.

[0025] Further, in step (2), when the photovoltaic system is running, the duty cycle D of the converter and the photovoltaic voltage U pv and bus voltage U bus The real-time sampling value of Input to the neural network, and obtain the identification value R of the equivalent loss resistance from the output of the neural network EQ *.

[0026] Furthermore, in step (3), the equivalent loss resistance R EQ *, by the photovoltaic voltage U pv , bus voltage U bus , the duty cycle D is used to observe the photovoltaic current according to the following formula:

[0027]

[0028] Where, T s is the switching period, L is the inductance value, I OB is the current observation value, k is the number of sampling cycles, and D' is the complementary value of the duty cycle, that is, D'=1-D.

[0029] The effects of the present invention are:

[0030] 1. By accurately identifying the equivalent loss resistance of the Boost DC converter, the online real-time and accurate observation of the photovoltaic current is realized, thus avoiding the problems of large loss, susceptibility to interference and high cost of the current sensor in the photovoltaic storage DC microgrid system;

[0031] 2. The present invention does not require accurate circuit component parameters or environmental sensors. It only requires voltage sampling circuits at the input and output sides of the converter, and uses sample data for offline training of neural networks to achieve environmentally adaptive photovoltaic current observation.

[0032] 3. The volume and economic cost of the photovoltaic system can be significantly reduced, which is conducive to further improving the comprehensive performance of the photovoltaic system. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the steps of the present invention

[0034] Figure 2 Boost DC converter circuit model (a) before considering equivalent loss resistance (b) after considering equivalent loss resistance

[0035] Figure 3 Elman neural network structure model

[0036] Figure 4 Simulation waveforms under different constant current conditions (a) Light intensity 1000W / m 2 Temperature 25℃(b) Light intensity 600W / m 2 Temperature 30℃

[0037] Figure 5 MPPT mode simulation waveform (a) Output connected to resistive load (b) Output connected to constant voltage source

[0038] Figure 6 Constant voltage control mode simulation waveform (light intensity 1000W / m 2 Temperature 25℃)

[0039] The main symbol names in the above figure: U pv , U bus are the photovoltaic voltage on the low voltage side and the DC bus voltage on the high voltage side respectively. high are the energy storage inductor and the high-voltage side capacitor respectively. L is the inductor current, R load is the output load. F is a freewheeling diode, Q1 and Q2 are power switch tubes. L , r DS , r C They are the parasitic resistances of the inductor, power switch, and high-voltage side capacitor, U F and r D is the diode conduction voltage drop and parasitic resistance. s , D are the switching period and duty cycle of the power switch tube, I L is the observed average inductor current. u(k), x(k), x c(k), y(k) are the outputs of the input layer, hidden layer, connecting layer and output layer of the kth iteration cycle respectively. x(k-1) is the output of the hidden layer of the k-1th iteration cycle. w1 is the weight coefficient from the input layer to the hidden layer, w2 is the weight coefficient from the connecting layer to the hidden layer, and w3 is the weight coefficient from the hidden layer to the output layer. b1 and b2 are the thresholds of the input layer and hidden layer respectively. DETAILED DESCRIPTION

[0040] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are described in further detail below.

[0041] Example 1

[0042] A specific implementation of the present invention is: a photovoltaic current observation method based on Elman neural network. Figure 1 The steps shown in the flowchart are Figure 2 The circuit model of the Boost DC converter in the embodiment of the present invention is specifically described with respect to the photovoltaic system parameters shown in the following table.

[0043] Step 1: Set the rated load resistance R on the output side of the Boost DC converter load =136Ω and half load (272Ω) to obtain the sample data set. During the data collection process, the duty cycle D of the power switch was changed between 0 and 0.75 under different light and temperature conditions, and the photovoltaic voltage U pv , bus voltage U bus Sampling is performed and D, U pv , U bus and R load Calculate the equivalent loss resistance R under each working condition according to the following formula EQ :

[0044]

[0045] Wherein D' is the complementary value of the duty cycle D, that is, D'=1-D.

[0046] And, by D, U pv , U bus and R load Calculate the slope of the IV output characteristic curve of the photovoltaic cell under various working conditions according to the following formula:

[0047]

[0048] Wherein, k is the number of sampling cycles, and D' is the complementary value of the duty cycle, that is, D'=1-D.

[0049] Table 1 PV system parameters

[0050]

[0051] The photovoltaic voltage U under each working condition is obtained pv , bus voltage U bus , the duty cycle D of the power switch and the slope of the IV output characteristic curve of the photovoltaic cell as a sample dataset.

[0052] Establish a four-input and one-output Elman neural network (ENN), whose structure is as follows Figure 3 As shown. The Elman neural network consists of an input layer, a hidden layer, a receiving layer, and an output layer. The input layer transmits input data to the hidden layer, and the output layer receives data from the hidden layer. The hidden layer and the output layer process the received data through the activation function. The receiving layer obtains data from the hidden layer and memorizes it, and inputs it into the hidden layer as feedback data for the next moment. The mathematical model of the Elman neural network is:

[0053] y(k)=g[w3x(k)+b2] (3)

[0054] x(k)=f[w1x c (k)+w2u(k)+b1] (4)

[0055]

[0056] Among them, u(k), x(k), x c (k) and y(k) are the outputs of the input layer, hidden layer, connecting layer and output layer of the kth period respectively. x(k-1) is the output of the hidden layer of the k-1th period. w1 is the weight coefficient from the input layer to the hidden layer, w2 is the weight coefficient from the connecting layer to the hidden layer, and w3 is the weight coefficient from the hidden layer to the output layer. b1 and b2 are the thresholds of the input layer and the hidden layer respectively.

[0057] In this embodiment, the number of hidden layers of the Elman neural network is 1, the input layer contains 4 neurons, and the output layer contains 1 neuron. The number of neurons in the hidden layer and the output layer is determined by the actual training effect. The input of the neural network is the photovoltaic voltage U pv , bus voltage U bus , the duty cycle D of the power switch and the slope of the IV output characteristic curve of the photovoltaic cell The output is the equivalent loss resistance R EQ .

[0058] Step 2: To prevent neuron saturation, the samples need to be normalized and the sample range is adjusted to (-1, 1) using the following formula.

[0059]

[0060] Where u x Represents the original input variable sample, u x * is the normalized input variable sample, u x_max and u x_min u x The number of neurons in the hidden layer is adjusted based on the following formula and further determined according to the actual training results.

[0061]

[0062] Where i, j, l are the number of neurons in the input layer, hidden layer, and output layer respectively, and h is an integer between 1 and 10.

[0063] The activation function f(x) of the hidden layer of the neural network uses the tansig function, whose output range is (-1,1) and is conducive to convergence. The activation function g(x) of the output layer uses a linear function. The activation function expressions of the hidden layer and output layer of the neural network are as follows:

[0064]

[0065] g(x)=x (9)

[0066] Finally, the fitting effect of the constructed Elman neural network was evaluated by the mean square error function (RMSE).

[0067]

[0068] Where N is the number of samples, n is the sample number, R n EQ_E is the estimated value of the neural network, R n EQ_S is the sample reference value.

[0069] The obtained sample set is used to train the Elman neural network model with different numbers of hidden layer neurons for multiple times. When the hidden layer contains 7 neurons, the training result with the smallest RMSE appears, and this training result is selected as the best neural network model. The obtained Elman neural network model is embedded in the digital controller to perform equivalent loss resistance R EQ Identification.

[0070] When the photovoltaic system is running, each switching cycle will pv , bus voltage U bus , the duty cycle D of the power switch and the slope of the IV output characteristic curve of the photovoltaic cell Input to the neural network, and obtain the identification value R of the equivalent loss resistance from the output of the neural network EQ *.

[0071] Step 3: Use the identified R EQ *, the following formula is used to conduct online real-time observation of photovoltaic current, thereby realizing efficient operation control of photovoltaic system.

[0072]

[0073] Where, T s is the switching period, L is the inductance value, I OB is the current observation value, k is the number of sampling cycles, and D' is the complementary value of the duty cycle, that is, D'=1-D.

[0074] In order to demonstrate the remarkable effects of the present invention, this embodiment provides some simulation results obtained by using the embodiments, and the implementation effects of the embodiments of the present invention are described below.

[0075] Figure 4 The following are simulation waveforms under different constant current conditions. When the constant current reference value increases step by step within the rated range, the photovoltaic current can always accurately follow the reference value, with a maximum error of only 13%. Moreover, even if the operating environment changes and the photovoltaic output characteristic curve changes, the photovoltaic system can still accurately observe the photovoltaic current through the neural network, without the need for additional environmental sensors, to achieve efficient control of the photovoltaic output current under different environments.

[0076] Figure 5 This is the MPPT mode simulation waveform. Figure 5 In (a), the output of the converter is connected to a resistive load. Under the action of the perturbation observation method, the photovoltaic operating point gradually approaches the maximum power operating point as the current perturbation occurs, achieving the iconic "three-stage" maximum power output steady-state operation. At this time, if the external environment changes, the photovoltaic maximum power point will shift. The photovoltaic operating point can be re-tracked and stabilized to the new maximum power point within 1 second. Figure 5 The output of the converter in (b) is connected to a 300V constant voltage source to simulate the operation of photovoltaic grid-connected and the working condition when the output side is connected to the energy storage battery. The simulation waveform shows that when the output side is constant voltage, the system can still accurately observe the photovoltaic current and achieve stable maximum power point tracking, which verifies the feasibility of the method of the present invention in an actual photovoltaic system.

[0077] Figure 6It is a simulation waveform of the constant voltage control mode. The corresponding environmental conditions at this time are a light intensity of 1000W / m2 and a temperature of 25°C. The reference value of the preset output voltage is 300V. Since the maximum output power of the photovoltaic cell in this environment is greater than the power required to maintain the reference voltage, the photovoltaic cell operates in the constant voltage mode to output a reduced power. Under the action of the DC bus voltage outer loop, the photovoltaic cell can stably output the DC bus voltage at a reference value of 300V based on the observed photovoltaic current. When the load switches between the rated 136Ω and half load, the overshoot and drop of the output voltage are both within 5%. During the dynamic response of the load step, the photovoltaic current can always be accurately observed. Therefore, the method of the present invention can meet the demand for stable DC bus voltage operation of the photovoltaic system.

[0078] The experimental results obtained in this embodiment can illustrate

[0079] 1. By accurately identifying the equivalent loss resistance of the Boost DC converter, the online real-time and accurate observation of the photovoltaic current is realized, thus avoiding the problems of high loss, susceptibility to interference and high cost of the current sensor in the photovoltaic storage DC microgrid system;

[0080] 2. The present invention does not require accurate circuit component parameters or environmental sensors. It only requires voltage sampling circuits at the input and output sides of the converter, and uses sample data for offline training of neural networks to achieve environmentally adaptive photovoltaic current observation.

[0081] 3. The volume and economic cost of the photovoltaic system can be significantly reduced, which is conducive to further improving the comprehensive performance of the photovoltaic system.

[0082] The present invention may be implemented in other specific forms without departing from its spirit or essential features. The described embodiments are considered to be illustrative and non-restrictive in all aspects, for example:

[0083] 1) Application topologies include various derivative topologies of Boost DC converters;

[0084] 2) Types and structures of neural networks;

[0085] 3) Selection of photovoltaic system parameters, etc.

[0086] The scope of the invention is therefore indicated by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalent technical solutions of the claims are embraced within their scope.

Claims

1. An environmentally adaptive photovoltaic current observation method based on a neural network, comprising the following steps: (1) Set a load resistor R of known resistance on the output side of the Boost DC converter load , calculate the equivalent loss resistance value R under the corresponding working conditions of each environment EQ , together with the voltage sampling data and duty cycle, constitute a sample set as the training data of the neural network; (2) Construct and train a neural network to convert the duty cycle D of the converter and the photovoltaic voltage U into pv and bus voltage U bus The real-time sampling value of Input to the neural network, and obtain the identification value R of the equivalent loss resistance from the output of the neural network EQ *; (3) Equivalent loss resistance R identified by neural network EQ *, by the photovoltaic voltage U pv , bus voltage U bus , the duty cycle D is used to observe the photovoltaic current according to the following formula: Where, T s is the switching period, L is the inductance value, I OB is the current observation value, k is the number of sampling cycles, D ’ is the complementary value of the duty cycle, that is, D ’ =1-D.

2. The method for observing photovoltaic current based on an environment-adaptive neural network according to claim 1, characterized in that: In step (1), the sample data set is obtained based on the actual circuit and voltage sampling. During the data acquisition process, the duty cycle of the power switch is changed, and the photovoltaic voltage U under each working condition is calculated. pv , bus voltage U bus Sampling values ​​for data collection.

3. The method for observing photovoltaic current based on an environment-adaptive neural network according to claim 1, characterized in that: In step (1), the sample data set of the neural network consists of the photovoltaic voltage U pv , bus voltage U bus , duty cycle D, slope of the IV output characteristic curve of the photovoltaic cell And the equivalent loss resistance value R of the converter EQ constitute.

4. The method for observing photovoltaic current based on an environment-adaptive neural network according to claim 3, characterized in that: Equivalent loss resistance R in the sample data set EQ The power switch duty cycle D and photovoltaic voltage U pv , bus voltage U bus and bus load value R load Calculation, R EQ The mathematical model is shown as follows: Where D ’ is the complementary value of the duty cycle, that is, D ’ =1-D.

5. The method for observing photovoltaic current based on an environment-adaptive neural network according to claim 3, characterized in that: The slope of the IV output characteristic curve of the photovoltaic cell in the sample data set The photovoltaic voltage U pv , bus voltage U bus , the duty cycle D is calculated according to the following formula: In the formula, k is the number of sampling cycles, D ’ is the complementary value of the duty cycle, that is, D ’ =1-D.

6. The method for observing photovoltaic current based on an environment-adaptive neural network according to claim 1, characterized in that: In step (2), a power switch duty cycle D, photovoltaic voltage U pv , bus voltage U bus , the slope of the IV output characteristic curve of the photovoltaic cell is the input, the equivalent loss resistance R EQ A neural network structure with four inputs and one output.

7. The method for observing photovoltaic current based on an environment-adaptive neural network according to claim 6, characterized in that: The specific internal structure of the neural network is determined by the actual training effect. Based on the sample data set obtained in the experiment, neural networks with different structures are trained multiple times. The training results are evaluated and compared using the error function, and the neural network with the best result is selected as the equivalent loss resistor R EQ The identification model is embedded into the digital controller.

Images