Piezoelectric energy collection system for maximum power point tracking of artificial neural network

Abstract

The invention discloses a piezoelectric energy collection system for maximum power point tracking of an artificial neural network, belongs to the field of micro energy collection, and realizes piezoelectric energy collection within a wide current source and large load change range. The system is composed of a piezoelectric element, a standard interface circuit experiment platform, an artificial neural network module, a standard interface circuit, a PID controller, a pulse width modulator, a DC-DC converter and a load. According to the invention, through combination of artificial neural network training, PID control and pulse width modulation, maximum power point tracking of the piezoelectric energy collection system is realized; by adjusting the size of the current source, the adaptive capacity of the maximum power point tracking system to the energy size change is verified; by adjusting the size of the load, the adaptive capacity of the maximum power point tracking system to the change of the size of the load is verified.

Description

Technical Field

[0001] The present application belongs to the field of micro energy collection, and specifically relates to a piezoelectric energy collection system for maximum power point tracking of an artificial neural network, which realizes piezoelectric energy collection within a wide current source and large load variation range. Background Art

[0002] Maximum power point tracking technology refers to finding the optimal voltage point that keeps the power at the maximum during the piezoelectric energy collection process, and adjusting the duty cycle of the DC-DC converter through a controller so that the piezoelectric energy collection system always operates at the maximum power point, thereby improving the efficiency of energy collection. The piezoelectric energy collection maximum power point tracking system based on artificial neural network proposed in this application has the same function as the traditional maximum power point tracking technology. The fundamental goal is to improve the efficiency of maximum power point tracking and make it more robust to current source and load changes. Compared with traditional maximum power point tracking technology, artificial neural network maximum power point tracking technology can better track changes in current source and load.

[0003] Wang proposed a maximum power point tracking technology that uses an envelope extractor to track the vibration amplitude. This method does not need to disconnect the piezoelectric energy harvester from the converter to measure the open-circuit voltage. It only directly extracts the output voltage and divides it through a voltage divider circuit composed of two resistors of the same resistance value to obtain V mpp. However, this method needs to be adjusted according to the environmental vibration conditions. (Wang, Xia Y, Shi G, et al. A novel MPPT techniquebased on the envelope extraction implemented with passive components forpiezoelectric energy harvesting[J]. IEEE Transactions on Power Electronics, 2021, 36(11): 12685-12693.). Fang et al. proposed a maximum power point tracking circuit based on inductor series synchronous switch acquisition. The system can measure the peak voltage during S-SSHI operation without open-circuit operation. The quality factor of the system is 2.68, but its adaptability to open-circuit voltage changes is poor. (Fang S, Xia H, Xia Y, et al. An efficient piezoelectric energy harvesting circuit with series-SSHI rectifier and FNOV-MPPT control technique[J]. IEEE Transactions on Industrial Electronics, 2020, 68(8): 7146-7155.). Gajendranath Chowdary proposed a modular power management system that can collect energy from three energy sources at the same time. The system uses a comparator-based open-circuit voltage method to achieve maximum power point tracking control, and uses an oscillator training algorithm to simulate the output characteristics of the comparator. The circuit is based on a 180 nm chip and can achieve 87% peak conversion efficiency and 1.5 V output voltage under the conditions of 20 µW input power and 0.6 V input voltage. (Chowdary G, Singh A, Chatterjee S. An18 nA, 87% efficient solar, vibration and RF energy-harvesting power management system with a single shared inductor[J]. IEEE Journal of solid-state circuits, 2016, 51(10): 2501-2513.). Summary of the invention

[0004] The present application provides a piezoelectric energy harvesting system with maximum power point tracking using an artificial neural network. The maximum power of the piezoelectric energy is predicted using an artificial neural network, and the duty cycle of the DC-DC converter is adjusted according to the predicted parameters to realize a piezoelectric energy harvesting maximum power point tracking system within a wide current source and large load variation range.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] On the one hand, the piezoelectric energy harvesting system for maximum power point tracking of an artificial neural network described in the present invention comprises: a piezoelectric element, a standard interface circuit experimental platform, an artificial neural network module, a standard interface circuit, a PID controller, a pulse width modulator, a DC-DC converter and a load;

[0007] The piezoelectric element and standard interface circuit experimental platform are used to obtain the voltage, current and power data required for artificial neural network training, and the piezoelectric element is used to simulate the output of piezoelectric vibration energy;

[0008] The artificial neural network module uses a multi-layer perceptron structure based on TensorFlow and Keras framework to predict the maximum power value of piezoelectric energy;

[0009] The standard interface circuit is used to collect the output of the piezoelectric equivalent model;

[0010] The PID controller is used to compare the optimal power predicted by the artificial neural network module with the actual power at the load end, calculate the duty cycle for output;

[0011] The pulse width modulator is used to output a pulse signal according to the duty cycle;

[0012] The DC-DC converter is used to adjust the voltage according to the pulse signal to achieve maximum power point tracking;

[0013] Preferably, the data preprocessing of the artificial neural network module includes feature normalization and rescaling of the target variable, and the scaling factor is determined by testing;

[0014] Preferably, the artificial neural network module includes an input layer, three hidden layers based on Relu activation function and an output layer for regression, and the hidden layer adopts early stopping and learning rate decay callback functions;

[0015] Preferably, the optimizer used by the neural network is Adam, and the loss function adopts mean square error, which penalizes prediction errors by squaring the error;

[0016] Preferably, the voltage and current values ​​at the rear end of the rectifier of the standard interface circuit experimental platform are used as input data of the artificial neural network module, and the load power at the rear end of the standard interface circuit experimental platform is used as output data of the artificial neural network;

[0017] Preferably, the standard interface circuit experimental platform comprises an oscilloscope, a power amplifier, a signal generator, a standard interface circuit and an excitation table, wherein the signal generator is connected to the power amplifier, the power amplifier is connected to the excitation table, the excitation table is connected to the standard interface circuit, and the oscilloscope is connected to the standard interface circuit;

[0018] Preferably, the standard interface circuit includes a piezoelectric equivalent model, a rectifier bridge, and a filter capacitor; the piezoelectric equivalent model is connected to the rectifier bridge; the filter capacitor is connected to the rectifier bridge;

[0019] Preferably, the piezoelectric equivalent model includes an equivalent current source, an internal capacitor and an internal resistor; the current source is connected to the internal capacitor; the internal resistor is connected to the internal capacitor;

[0020] Preferably, in the piezoelectric energy harvesting system for maximum power point tracking using an artificial neural network, the standard interface circuit is connected to the DC-DC converter, and the DC-DC converter is connected to the load; the artificial neural network module is connected to the PID controller; the PID controller is connected to the pulse width modulator, and the duty cycle is calculated and output by comparing the optimal power predicted by the artificial neural network with the actual power at the load end; the pulse width modulator is connected to the DC-DC converter, and is used to output a pulse signal according to the duty cycle; the DC-DC converter is used to adjust the voltage according to the pulse signal to achieve maximum power point tracking;

[0021] Through the above technical solutions, it can be known that compared with the prior art, the present application discloses a piezoelectric energy harvesting system with maximum power point tracking of an artificial neural network, which predicts the maximum power of piezoelectric energy by using an artificial neural network, adjusts the duty cycle of the DC-DC converter according to the predicted parameters, and realizes the maximum power point tracking of piezoelectric energy harvesting within a wide current source and a large load variation range. The preprocessing adopted by the present invention includes feature normalization and rescaling of the target variable, and the scaling factor is determined by testing to improve the accuracy of the model; the present invention adopts a ReLU activation function, which enables the neural network to effectively learn the complex nonlinear relationship in the piezoelectric energy harvesting system through its nonlinear characteristics; the neural network contains an output layer for regression, which is used to output continuous values ​​and directly predict the optimal power value of the system. The layer adopts callback functions such as early stopping and learning rate decay to prevent overfitting and improve generalization ability. The optimizer used by the neural network is Adam, which is an optimization algorithm with an adaptive learning rate. The loss function is the mean square error, which calculates the average of the squared difference between the predicted value and the actual value. The mean square error penalizes prediction errors by squaring the error, thereby amplifying the impact of outliers. The present invention adopts artificial neural network to realize maximum power point tracking of piezoelectric energy collection. The system can maintain good maximum power point tracking efficiency within a wide current source and large load variation range. The piezoelectric energy collection system has good robustness to current source and load variations. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solution of the embodiment of the present invention, the following will describe the

[0023] The accompanying drawings are used for a brief introduction. Obviously, the accompanying drawings in the following description are only embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on the provided accompanying drawings without paying any creative work.

[0024] Figure 1 :Schematic diagram of the structure of the maximum power point tracking system for piezoelectric energy harvesting based on artificial neural network;

[0025] Figure 2 : Standard interface circuit experimental system diagram;

[0026] Figure 3 : Structural diagram of artificial neural network model;

[0027] Figure 4 : Schematic diagram of artificial neural network training loss function;

[0028] Figure 5 : Schematic diagram of the comparison between the maximum power point tracking system based on artificial neural network training and the load power of the standard circuit;

[0029] Figure 6 : Output power waveform of the system under different current sources;

[0030] Figure 7 :The relationship between the output power of the proposed system and the current source;

[0031] Figure 8 :Maximum power point tracking efficiency curve of piezoelectric energy harvesting system when current source changes;

[0032] Figure 9 : The proposed system output power changes with load;

[0033] Figure 10 : Maximum power point tracking efficiency curve of piezoelectric energy harvesting system when load changes; DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] As described in the background art, it is difficult for the prior art to stably implement maximum power point tracking of piezoelectric energy harvesting under environmental changes.

[0036] To solve the above problems, an embodiment of the present application proposes a piezoelectric energy harvesting system with maximum power point tracking of an artificial neural network. By combining model optimization training of the artificial neural network with PID control, maximum power point tracking of piezoelectric energy harvesting within a wide current source and large load variation range is achieved.

[0037] The structure of the piezoelectric energy harvesting system for maximum power point tracking of an artificial neural network described in the present invention is as follows: Figure 1As shown, the energy harvesting experimental system is composed of a piezoelectric element, a standard interface circuit, a DC-DC converter, a pulse width modulator, a PID controller, an artificial neural network module and a load, wherein the pulse width modulator and the PID controller form a maximum power point tracking controller. The energy harvesting experimental system composed of the piezoelectric element and the standard interface circuit measures the voltage, current and output power, and uses the measured voltage and current data as the input of the artificial neural network training, and the measured output power as the output of the artificial neural network. After training, the artificial neural network predicts the maximum power point of the system. The PID controller simultaneously receives the power predicted by the artificial neural network and the actual power at the load end, and calculates the duty cycle output according to the error between the predicted power and the actual power. The pulse width modulator adjusts the DC-DC converter according to the duty cycle output by the controller, thereby realizing maximum power point tracking.

[0038] In order to obtain the input and output data of artificial neural network, a standard interface circuit experimental platform was established, such as Figure 2 The experimental platform includes an oscilloscope, a power amplifier, a signal generator, a standard interface circuit and an excitation table. The voltage and current output from the rear end of the rectifier in the standard circuit are used as the input of the artificial neural network, and the load resistance power is used as the output.

[0039] The structure diagram of the artificial neural network model is as follows Figure 3 As shown. Based on the acquired data set, the artificial neural network is trained using Python. The training process uses a multi-layer perceptron structure based on the TensorFlow and Keras framework. The artificial neural network is designed to predict the power on the load side based on the input voltage and current data. Since the data values ​​are small and the changes between adjacent data points are small, the data needs to be preprocessed to improve the model performance. The preprocessing includes feature normalization and rescaling of the target variable. The scaling factor is determined by testing and is set to 1000.

[0040] During the training process, the three hidden layers all use the ReLU activation function, which enables the neural network to effectively learn the complex nonlinear relationships in the piezoelectric energy harvesting system through its nonlinear characteristics. The first hidden layer has 128 neurons and the second hidden layer has 64 neurons, which captures more feature information by increasing the complexity of the network. The third hidden layer contains 32 neurons, and the number of neurons is gradually reduced to capture higher-level abstract features. The neural network contains an output layer for regression, which is used to output continuous values ​​and directly predict the optimal power value of the system. This layer uses callback functions such as early stopping and learning rate decay to prevent overfitting and improve generalization ability. The performance of the model is evaluated on the test set to obtain the validation loss, and the prediction results are converted back to the original scale. The optimizer used by the neural network is Adam, which is an optimization algorithm with an adaptive learning rate. The loss function is the mean square error, which calculates the average of the squared differences between the predicted value and the actual value. The mean square error penalizes prediction errors by squaring the error, thereby amplifying the impact of outliers. The loss function obtained during training is shown in the figure below. Figure 4 After the training process, both the training loss and the validation loss showed a trend of continuous decline and convergence, and finally stabilized below 0.1, which shows that the model effectively captured the data features during the learning process and did not overfit.

[0041] Build a simulation environment in Matlab / Simulink, and select the current source I p 5mA, 20Hz, clamping capacitor C p is 50nF, internal resistance R p The filter capacitor C is 10kΩ. r The load resistor is 20μF, the inductor L is 1H, the capacitor C1 is 33nF, and the optimal load is 100kΩ (the resistance value is the same as the resistance value used in the experiment). The predicted power and actual power are input into the PID controller to measure the error, which is used to adjust the duty cycle of the DC-DC converter and then converted into a control signal through the pulse width modulator. The actual power at the load end is obtained from the measured data simulated in Simulink. The input and output of the DC-DC can be expressed by formula (1):

[0042]

[0043] Where T ON is the switch on time, T OFF is the switch off time. The duty cycle expression is:

[0044] (2)

[0045] When the current source is set to 5mA and the load resistance is the optimal load of 100kΩ, Figure 5 This is a schematic diagram comparing the maximum power point tracking system based on artificial neural network training with the standard circuit load power. As shown in the figure, after 2.5 seconds, the system has reached the maximum power point, and the average output power has increased from 10.23mW to 11.76mW. At the same time, the oscillation of the load resistance power waveform has been greatly reduced. When the power drops to the lowest point (1 second), the power has increased by 57.1%, which shows that the system collects energy more stably. At this time 11.76mW, The maximum power point tracking efficiency is 96.5%. The maximum power point tracking efficiency is calculated using formula (3):

[0046] (3)

[0047] In the formula, represents the average power at the load end after the proposed maximum power point tracking system, Indicates the maximum power at the load end of the standard interface circuit.

[0048] By changing the current source, the output power waveform of the proposed system is shown in Figure 6 As shown. At this time, the calculation of the maximum power point tracking efficiency is carried out when the load resistance is kept at the optimal load resistance of 100kΩ. As shown in the figure, the system achieves maximum power point tracking when the current source is 0.7mA, 1mA, and 3mA, and the tracking efficiency is 95.7%, 95.9%, and 96%, respectively. When Ip=5mA and Rload=100kΩ, the maximum power point tracking efficiency reaches its peak, and the peak efficiency is 96.5%. When the current source is 0.7mA and 1mA, the system takes about 0.8 seconds to achieve maximum power point tracking; when the current source is 3mA, the system takes about 1 second to achieve maximum power point tracking; and when the current source is 5mA, it takes about 1.5 seconds to achieve maximum power point tracking. This shows that even if the external vibration conditions change and the energy is different, the proposed system can still efficiently collect energy at the maximum power point.

[0049] Figure 7The following is a graph showing the relationship between the output power of the proposed system and the change in the current source. When the parameters are set to Ip=5mA and Rload=100kΩ, the system achieves a peak efficiency of 96.5%. When the current source changes, the system can still achieve maximum power point tracking and maintain a high efficiency. In order to verify the adaptability of the present invention to changes in input energy, under the premise that the load remains unchanged, the current source is set to 0.2mA, 0.3mA, 0.5mA, 0.7mA, 1mA, 3mA, 5mA (parameters used for artificial neural network training), 10mA, 12mA and 13mA, respectively. The maximum power point tracking efficiencies obtained are 93.2%, 94.6%, 95.8%, 96.5% (parameters used for artificial neural network training), 95.7%, 95.9%, 96%, 94.6%, 93.6%, 92.7%, respectively. Figure 8 As shown in the figure, the maximum tracking efficiency is 96.5%, and when the current source changes from 0.2mA to 13mA, the maximum power point tracking efficiency exceeds 92.7%. This verifies that the present invention can still achieve maximum power point tracking of piezoelectric energy harvesting when the energy in the environment changes.

[0050] Figure 9 The output power of the proposed system varies with the load. The points where the maximum power point tracking efficiency exceeds 96% are marked with circles. As shown in the figure, as the load resistance changes from 80KΩ to 10MΩ, the maximum power point tracking efficiency remains above 92.3%, which shows that the proposed system can adapt to changes in load resistance. In order to verify the adaptability of the present invention to load changes, under the premise that the current source remains unchanged, the load resistance is set to 60 kΩ, 80 kΩ, 100 kΩ (parameters used for artificial neural network training), 120 kΩ, 140 kΩ, 160 kΩ, 180 kΩ, 200 kΩ, 400 kΩ, 1000 kΩ, 5000 kΩ and 10000 kΩ, and the corresponding maximum power point tracking efficiencies are 86.6%, 92.3%, 96.5% (parameters used for artificial neural network training), 96.1%, 96.4%, 95.5%, 95.4%, 96.2%, 96.1%, 95.8%, 95.4% and 88%, respectively. Figure 10 When the load resistance changes in the range of 100 kΩ to 5 MΩ, the maximum power point tracking efficiency exceeds 95%, which verifies that the present invention can still achieve maximum power point tracking of piezoelectric energy harvesting when the load changes.

[0051] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A piezoelectric energy harvesting system for maximum power point tracking of an artificial neural network, characterized in that: It consists of piezoelectric element, standard interface circuit experimental platform, artificial neural network module, standard interface circuit, PID controller, pulse width modulator, DC-DC converter and load; The piezoelectric element and standard interface circuit experimental platform are used to obtain the voltage, current and power data required for artificial neural network training, and the piezoelectric element is used to simulate the output of piezoelectric vibration energy; The artificial neural network module predicts the maximum power value of piezoelectric energy based on the multi-layer perceptron structure of TensorFlow and Keras framework; The standard interface circuit is used to collect the output of the piezoelectric equivalent model; The PID controller is used to compare the optimal power predicted by the artificial neural network with the actual power at the load end, calculate the duty cycle for output; The pulse width modulator is used to output a pulse signal according to the duty cycle; The DC-DC converter is used to adjust the voltage according to the pulse signal to achieve maximum power point tracking.

2. The piezoelectric energy harvesting system for maximum power point tracking of an artificial neural network as claimed in claim 1, characterized in that: The standard interface circuit experimental platform comprises an oscilloscope, a power amplifier, a signal generator, a standard interface circuit and an excitation table; the signal generator is connected to the power amplifier, the power amplifier is connected to the excitation table, the excitation table is connected to the standard interface circuit, and the oscilloscope is connected to the standard interface circuit.

3. The piezoelectric energy harvesting system for maximum power point tracking of an artificial neural network as claimed in claim 1, characterized in that: Data preprocessing for the ANN module includes feature normalization and rescaling of the target variable, with the scaling factor determined through testing; The artificial neural network module uses a multi-layer perceptron structure based on TensorFlow and Keras framework; the artificial neural network module includes an input layer, three hidden layers based on Relu activation function and an output layer for regression, and the hidden layer adopts early stopping and learning rate decay callback functions; the optimizer used by the artificial neural network module is Adam, and the loss function adopts mean square error, which punishes prediction errors by squaring the error.

4. The piezoelectric energy harvesting system for maximum power point tracking of an artificial neural network as claimed in claim 1, characterized in that: The standard interface circuit comprises a piezoelectric equivalent model, a rectifier bridge and a filter capacitor; the piezoelectric equivalent model is connected to the rectifier bridge; and the filter capacitor is connected to the rectifier bridge.

5. The piezoelectric energy harvesting system for maximum power point tracking of an artificial neural network as claimed in claim 1, characterized in that: The standard interface circuit is connected to the DC-DC converter, and the DC-DC converter is connected to the load; the artificial neural network module is connected to the PID controller; the PID controller is connected to the pulse width modulator; and the pulse width modulator is connected to the DC-DC converter.

6. The piezoelectric energy harvesting system for maximum power point tracking of an artificial neural network as claimed in claim 4, characterized in that: The piezoelectric equivalent model includes an equivalent current source, an internal capacitor and an internal resistor; the current source is connected to the internal capacitor; and the internal resistor is connected to the internal capacitor.

Images