Intelligent optical transformer and method of using the same

Abstract

The application discloses an intelligent photoelectric transformer and a use method thereof, and belongs to the technical field of transformers. The transformer comprises a photoelectric acquisition unit, a signal conditioning and analog-digital conversion module, a data processing and core control module, a multi-protocol data communication module and a power management module, the photoelectric acquisition unit is connected with the signal conditioning and analog-digital conversion module, the signal conditioning and analog-digital conversion module is connected with the data processing and core control module, and the data processing and core control module is connected with the multi-protocol data communication module. The application deeply integrates photoelectric sensing, variable step adaptive filtering and multi-scale time domain convolution network, realizes high-precision electric energy metering, multi-dimensional electric energy quality analysis and load trend prediction, supports double-protocol communication, has strong real-time processing capability and high compatibility, upgrades a traditional sensor into a sensing terminal with edge intelligence, is a core equipment supporting digital transformation of a smart power grid, and is convenient to use.

Description

Technical Field

[0001] This invention relates to the field of current transformer technology, and specifically to an intelligent photoelectric current transformer and its usage method. Background Technology

[0002] Power transformers are core equipment in substations for measuring and protecting electrical quantities, and their performance directly affects the safe operation of the power grid and the accuracy of metering. Traditional electromagnetic transformers, based on the principle of electromagnetic induction, suffer from problems such as magnetic saturation and ferroresonance, have a narrow dynamic range and poor frequency response, and cannot meet the requirements of modern power grids for measuring harmonics, transients, and other non-power frequency components. Therefore, electronic transformers (such as Rogowski coils and optical transformers) are gradually being adopted. They achieve high-low voltage isolation through photoelectric sensing and output digital signals, overcoming the shortcomings of electromagnetic transformers to some extent.

[0003] However, most existing electronic instrument transformers only output raw sampled values ​​or fundamental frequency data after simple digital filtering, lacking the ability to deeply analyze the power grid's operating status. Their internal signal processing typically uses fixed-parameter filters, which become less effective when the power grid frequency fluctuates or noise characteristics change over time. Furthermore, their data output format is limited, failing to provide advanced information such as power quality and load trends simultaneously, making it difficult to meet the needs of smart grids for ubiquitous data sensing and edge computing.

[0004] In recent years, deep learning has shown potential in power system analysis, with some studies attempting to apply recurrent neural networks or convolutional neural networks for power quality classification or load forecasting. However, existing models are mostly designed for single tasks and use single-scale convolution, failing to effectively capture the multi-scale characteristics of millisecond-level harmonic transients and minute-level load trends in grid signals; the models also have high computational demands, making them difficult to deploy in embedded devices with high real-time requirements. Furthermore, existing instrument transformer communication interfaces mostly support a single protocol (such as IEC 61850 or Modbus), requiring additional protocol conversion when interconnecting with equipment from different manufacturers, increasing system complexity and cost.

[0005] To address the aforementioned issues, there is an urgent need for an intelligent photoelectric transformer and its application method to resolve the problems associated with traditional methods. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent photoelectric transformer and its usage method, which deeply integrates photoelectric sensing, variable step size adaptive filtering and multi-scale time-domain convolutional network to achieve high-precision power metering, multi-dimensional power quality analysis and load trend prediction. It also supports dual-protocol communication, has strong real-time processing capabilities and high compatibility, and upgrades traditional sensors into sensing terminals with edge intelligence. It is a core device supporting the digital transformation of smart grids and is easy to use.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An intelligent photoelectric transformer includes: a photoelectric acquisition unit, a signal conditioning and analog-to-digital conversion module, a data processing and core control module, a multi-protocol data communication module, and a power management module. The photoelectric acquisition unit is connected to the signal conditioning and analog-to-digital conversion module, the signal conditioning and analog-to-digital conversion module is connected to the data processing and core control module, and the data processing and core control module is connected to the multi-protocol data communication module. The photoelectric acquisition unit is used to sense the current in the primary conductive path and convert it into an optical signal output; The signal conditioning and analog-to-digital conversion module is used to receive optical signals and convert them into discrete digital signal sequences; The data processing and core control module includes an active filtering unit and a temporal convolutional network state analysis unit. The active filtering unit uses a variable step-size normalized least mean square adaptive filter to filter discrete digital signal sequences. The variable step-size normalized least mean square adaptive filter dynamically adjusts the step-size factor based on the energy of the error signal to output a clean signal, where the error signal is the difference between the expected response and the actual output of the filter. The temporal convolutional network state analysis unit extracts features from the clean signal based on a multi-scale dilated temporal convolutional network. The multi-scale dilated temporal convolutional network includes at least three parallel branches, each branch using stacked dilated causal convolutional layers with different dilation factor combinations. The output features of each branch are concatenated and processed by an attention enhancement module, and then output through three parallel fully connected output heads to output power metering data, power quality monitoring indicators, and load trend analysis results, respectively. The multi-protocol data communication module is used to encapsulate power metering data, power quality monitoring indicators and load trend analysis results according to IEC 61850-9-2 protocol and Modbus TCP / IP protocol respectively, and then send them to the substation bay layer or station control layer equipment through the fiber optic Ethernet interface. The power management module is used to provide operating voltage for each module.

[0008] Furthermore, the photoelectric acquisition unit includes: a Rogowski coil, an integrator, an electro-optical conversion circuit, and an optical fiber. The Rogowski coil is installed on the primary high-voltage bus and connected to the integrator. The integrator is connected to the electro-optical conversion circuit. The electro-optical conversion circuit is connected to the signal conditioning and analog-to-digital conversion module through the optical fiber. The Rogowski coil is used to sense the current being measured and generate an induced electromotive force. The integrator is used to integrate the induced electromotive force to obtain a voltage signal; The electro-optic conversion circuit is used to linearly modulate a voltage signal into an optical signal; The optical fiber is used to transmit optical signals to the signal conditioning and analog-to-digital conversion module.

[0009] Furthermore, the signal conditioning and analog-to-digital conversion module includes: a photodetector, a transimpedance amplifier, an anti-aliasing filter, an analog-to-digital converter, and a digital interface circuit. The optical fiber is connected to the photodetector, the photodetector is connected to the transimpedance amplifier, the transimpedance amplifier is connected to the anti-aliasing filter, the anti-aliasing filter is connected to the analog-to-digital converter, the analog-to-digital converter is connected to the digital interface circuit, and the digital interface circuit is connected to the data processing and core control module. The photodetector is used to convert the received optical signal into photocurrent; The transimpedance amplifier is used to convert photocurrent into a voltage signal; The anti-aliasing filter is a second-order active low-pass filter; The analog-to-digital converter is used to sample and quantize the input voltage and output a 16-bit binary two's complement digital code value. The digital interface circuit is used to transmit 16-bit binary two's complement digital code values ​​to the data processing and core control module in real time via a serial peripheral interface bus.

[0010] Furthermore, the energy metering data includes the fundamental voltage amplitude, fundamental current amplitude, phase difference, fundamental frequency, and active power calculated accordingly.

[0011] Furthermore, the power quality monitoring indicators include total harmonic distortion rate and the voltage amplitude of each harmonic.

[0012] Furthermore, the load trend analysis results include predicted load current values ​​for future sampling points.

[0013] The present invention also provides a method for using an intelligent photoelectric transformer, applied to the aforementioned intelligent photoelectric transformer, comprising: Step 1: The photoelectric acquisition unit senses the primary side current and converts it into an optical signal. The signal conditioning and analog-to-digital conversion module receives the optical signal and converts it into a digital signal sequence. Step 2: The active filtering unit uses a variable step-size normalized least mean square adaptive filter to filter the digital signal sequence. The step-size factor is dynamically adjusted based on the energy of the error signal to output a clean signal. Step 3: The temporal convolutional network state analysis unit extracts features from the clean signal based on a multi-scale dilated temporal convolutional network, and outputs power metering data, power quality monitoring indicators and load trend analysis results in parallel. Step 4: The multi-protocol data communication module encapsulates the results according to the IEC 61850-9-2 protocol and the Modbus TCP / IP protocol respectively, and then sends them to the substation bay layer or station control layer equipment through the fiber optic Ethernet interface.

[0014] Furthermore, in step 2, the active filtering unit uses a variable step-size normalized least mean square adaptive filter to filter the digital signal sequence, dynamically adjusting the step-size factor based on the energy of the error signal to output a clean signal, specifically: Construct a transverse finite impulse response filter and initialize the weight vector to zero; The desired response is generated by estimating the grid frequency and phase in real time using a digital phase-locked loop. Calculate the output and error of the transverse finite impulse response filter; Calculate the energy estimate of the error signal; The step size factor is dynamically calculated based on the energy estimation of the error signal; The weight vector is updated using the normalized least mean square algorithm; Repeat the above process until convergence is achieved, resulting in a filtered, clean signal.

[0015] Furthermore, in step 3, the temporal convolutional network state analysis unit extracts features from the clean signal based on a multi-scale dilated temporal convolutional network, and outputs in parallel power metering data, power quality monitoring indicators, and load trend analysis results, specifically as follows: Extract the input tensor from the clean signal according to the time window; Dilated causal convolution is performed on the three parallel branches of the multi-scale feature extraction module to obtain the output features of the three branches, and then the multi-scale fused features are concatenated in the channel dimension. Enhanced features are obtained by performing channel attention and spatial attention enhancement on multi-scale fusion features using an attention feature enhancement module. Global average pooling is used to obtain the feature vector of the enhanced features; The feature vectors are fed into three parallel fully connected output heads to obtain power metering data, power quality monitoring indicators, and load trend analysis results.

[0016] Furthermore, in step 4, the multi-protocol data communication module encapsulates the obtained results according to the IEC 61850-9-2 protocol and the Modbus TCP / IP protocol respectively, and then sends them to the substation bay layer or station control layer equipment through the fiber optic Ethernet interface, specifically: The data set of electricity metering data, power quality monitoring indicators and load trend analysis results is constructed according to the IEC 61850-9-2 protocol, encapsulated in Ethernet frames after ASN.1 encoding, and sent to the process layer network through the first optical module in a periodic or event-triggered manner. According to the predefined Modbus register address mapping table, the data is stored in the corresponding register. The second optical module responds to the Modbus TCP / IP read request from the station control layer network, encapsulates the data, and sends it.

[0017] In summary, the present invention has at least one of the following beneficial technical effects: 1. This invention solves the problems of existing instrument transformers having limited functionality and low intelligence in output data: by integrating active filtering and deep learning models, the traditional instrument transformer is upgraded from a passive signal sensor to an active sensing terminal with edge intelligence, directly outputting multi-dimensional information such as power metering, quality monitoring, and load forecasting.

[0018] 2. This invention solves the problem of the contradiction between convergence speed and steady-state accuracy caused by the fixed step size of traditional adaptive filters: the proposed variable step size algorithm can dynamically adjust the step size according to the error energy, which can not only track quickly when the power grid is disturbed, but also maintain high accuracy in steady state, effectively suppressing photoelectric noise and electromagnetic interference.

[0019] 3. This invention solves the problem that single-scale deep learning models have difficulty capturing the multi-timescale features of power grid signals: the multi-scale dilated convolutional structure enables the network to perceive the temporal dependencies of different scales at the same time, and the attention mechanism further strengthens key features, significantly improving the accuracy of tasks such as harmonic analysis and load forecasting.

[0020] 4. This invention solves the problems of resource waste and insufficient real-time performance caused by multi-task independent modeling: multi-tasks share the feature extraction layer, output the results of three tasks in parallel, the inference time is less than 60μs, which meets the real-time processing requirements of 12.8kHz sampling rate, and can be deployed on embedded platforms.

[0021] 5. This invention solves the problem of poor equipment compatibility caused by inconsistent communication protocols: Dual protocol support enables this invention to be seamlessly integrated into substation automation systems of different manufacturers and standards, improving equipment interoperability and reducing system integration costs. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the intelligent photoelectric transformer structure of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0024] like Figure 1 As shown, the present invention provides an intelligent photoelectric transformer, comprising: a photoelectric acquisition unit, a signal conditioning and analog-to-digital conversion module, a data processing and core control module, a multi-protocol data communication module, and a power management module. The photoelectric acquisition unit is connected to the signal conditioning and analog-to-digital conversion module, the signal conditioning and analog-to-digital conversion module is connected to the data processing and core control module, and the data processing and core control module is connected to the multi-protocol data communication module. The photoelectric acquisition unit is used to sense the current in the primary conductive path and convert it into an optical signal output; The signal conditioning and analog-to-digital conversion module is used to receive optical signals and convert them into discrete digital signal sequences; The data processing and core control module includes an active filtering unit and a temporal convolutional network state analysis unit. The active filtering unit uses a variable step-size normalized least mean square adaptive filter to filter discrete digital signal sequences. The variable step-size normalized least mean square adaptive filter dynamically adjusts the step-size factor based on the energy of the error signal to output a clean signal, where the error signal is the difference between the expected response and the actual output of the filter. The temporal convolutional network state analysis unit extracts features from the clean signal based on a multi-scale dilated temporal convolutional network. The multi-scale dilated temporal convolutional network includes at least three parallel branches, each branch using stacked dilated causal convolutional layers with different dilation factor combinations. The output features of each branch are concatenated and processed by an attention enhancement module, and then output through three parallel fully connected output heads to output power metering data, power quality monitoring indicators, and load trend analysis results, respectively. The multi-protocol data communication module is used to encapsulate power metering data, power quality monitoring indicators and load trend analysis results according to IEC 61850-9-2 protocol and Modbus TCP / IP protocol respectively, and then send them to the substation bay layer or station control layer equipment through the fiber optic Ethernet interface. The power management module is used to provide operating voltage for each module.

[0025] The photoelectric acquisition unit includes: a Rogowski coil, an integrator, an electro-optical conversion circuit, and an optical fiber. The Rogowski coil is located on the primary high-voltage bus and connected to the integrator. The integrator is connected to the electro-optical conversion circuit. The electro-optical conversion circuit is connected to the signal conditioning and analog-to-digital conversion module via an optical fiber. The sensing head of the Rogowski coil is mounted on the high-voltage busbar being measured. It consists of a non-magnetic frame and a uniformly wound coil, and the coil outputs an induced electromotive force at both ends. When the measured current flows through the busbar At that time, according to the law of electromagnetic induction, the induced electromotive force is directly proportional to the rate of change of the measured current: ; in, is the mutual inductance coefficient of the coil, measured in Henry (H), and is determined by the coil's geometric parameters and number of turns. Because It is proportional to the derivative of the current; in order to restore the current waveform, an integrator needs to be connected.

[0026] The integrator employs a high-precision analog integrator circuit, consisting of an operational amplifier and resistors. and capacitor C Composition, on Perform integration and output the sum. Proportional voltage signal : ; in, '、 ' is the integration time constant, The initial voltage for integration is typically reset to zero by a reset circuit before each measurement. Integrator output. It is fed into the electro-optical conversion circuit.

[0027] The core of the electro-optical conversion circuit is a semiconductor laser diode, whose output optical power is proportional to the driving current. The laser diode's luminous intensity is linearly modulated after passing through a voltage-to-current conversion circuit, thereby increasing the optical power of the output optical signal. Follow Linear change: ; in, The electro-optic conversion coefficient is expressed in W / V. To bias the optical power, the laser diode is ensured to operate in the linear region. The modulated optical signal is coupled into the optical fiber through a lens and transmitted through an insulated fiber optic column to the signal conditioning and analog-to-digital conversion module located on the low-voltage side, achieving electrical isolation between the primary and secondary sides.

[0028] The signal conditioning and analog-to-digital conversion module receives optical signals from optical fibers, converts them back into electrical signals, and then conditions and converts them into discrete digital signal sequences. The module internally comprises, in sequence, a photodetector, a transimpedance amplifier, an anti-aliasing filter, an analog-to-digital converter (ADC), and a digital interface circuit. The optical fiber connects to the photodetector, the photodetector connects to the transimpedance amplifier, the transimpedance amplifier connects to the anti-aliasing filter, the anti-aliasing filter connects to the ADC, the ADC connects to the digital interface circuit, and the digital interface circuit connects to the data processing and core control module. The photodetector uses a PIN photodiode, and its responsivity It is constant at a given light wavelength. When the received light power... When the PIN diode is illuminated, a weak photocurrent proportional to its size is generated. : ; The unit is A / W. This photocurrent is fed into a transimpedance amplifier, which converts the current signal into a voltage signal. The conversion relationship is determined by the feedback resistor. Decide: ; A second-order active low-pass filter is connected after the transimpedance amplifier for anti-aliasing filtering. The cutoff frequency of this filter is... Designed for analog-to-digital converter sampling frequency Half of, that is This is done to eliminate noise and interference above the Nyquist frequency and prevent spectral aliasing after sampling. The filter output is a fully smoothed analog voltage signal. .

[0029] The analog-to-digital converter (ADC) employs a high-precision successive approximation ADC with a 16-bit resolution and a bipolar input range. Its sampling clock is provided by a high-stability temperature-compensated crystal oscillator, with the frequency precisely set to [frequency value missing]. This ensures that 256 points are sampled within each power frequency cycle (20ms). At each sampling time... ( (sampling interval), analog-to-digital converter pair Quantization is performed to output a 16-bit binary two's complement digital code value. Its relationship with the input voltage is as follows: ; in, This is the reference voltage for the analog-to-digital converter, taken as 10V; This indicates integer quantization. This is the final digital signal sequence, with a dimension of 1, representing the sampling time. The digital code corresponding to the instantaneous value.

[0030] The analog-to-digital converter transmits data via a serial peripheral interface bus. Data is transmitted in real time to the data processing and core control module. The serial peripheral interface clock frequency is set to 10MHz, and the data transmission format is 16 bits, with the most significant bit first. The field-programmable gate array (FPGA) section in the data processing and core control module acts as the serial peripheral interface host. Upon each sampling interrupt, it reads the conversion result from the analog-to-digital converter and stores it in a dual-port random access memory for subsequent use by the active filtering unit.

[0031] The hardware carrier of the data processing and core control module is a heterogeneous computing chip that uses a field-programmable gate array (FPGA) and a digital signal processor (DSP) for collaborative processing. The active filtering unit and the temporal convolutional network state analysis unit will be explained in detail in the subsequent methods.

[0032] The multi-protocol data communication module receives structured data frames output by the data processing and core control module, encapsulates them according to multiple protocols, and transmits them through a fiber optic Ethernet interface. The hardware includes a field-programmable gate array (FPGA) internal media access control layer soft core, a gigabit physical layer transceiver, two small pluggable optical modules, and a real-time clock chip.

[0033] The Media Access Control (MAC) layer soft core supports IEEE 1588 Precision Time Protocol hardware timestamps, synchronized with the sampling time in step 1. The physical layer transceiver connects to the MACC layer via a simplified gigabit media-independent interface, driving a small, pluggable optical module (single-mode 1310nm, 10km transmission distance). A real-time clock chip provides high-precision time.

[0034] Protocol stack implementation: The IEC 61850-9-2 sampling value protocol is used to transmit power metering and harmonic data, employing multicast Ethernet frames and encapsulating ASN.1 encoded report messages; Modbus TCP / IP is used to transmit power quality and load forecast data, employing a TCP / IP protocol stack with predefined register address mappings. The two ports are connected to the process layer network and the station control layer network, respectively.

[0035] Data encapsulation process: After each inference operation, the data processing module writes the data to shared memory and triggers an interrupt. The communication module reads the data and encapsulates it according to a preset protocol: For IEC 61850-9-2, a dataset containing all output data is constructed, modeled according to IEC 61850 logical nodes, and encapsulated in an Ethernet frame after ASN.1 encoding; for Modbus TCP / IP, the data is filled into the response message according to register address mapping. Port 1 uses periodic (100ms) and event-triggered transmission, while port 2 acts as a TCP server responding to client read requests.

[0036] All transmitted frames are accompanied by an IEEE 1588 hardware timestamp to ensure accurate time reconstruction at the receiving end. The communication module acts as a precision time protocol slave clock, synchronized with the station control layer master clock, with a synchronization accuracy better than 100ns.

[0037] The electricity metering data includes the fundamental voltage amplitude, fundamental current amplitude, phase difference, fundamental frequency, and active power calculated accordingly.

[0038] The power quality monitoring indicators include total harmonic distortion rate and voltage amplitude of each harmonic.

[0039] The load trend analysis results include predicted load current values ​​for future sampling points.

[0040] The power management module provides stable and isolated operating voltages for each module. It employs a dual-redundant DC / DC converter design, drawing power from the station's DC power supply (typical input voltage 110V or 220V DC). The first output, +5V / 10A, is the main power supply, which is converted to +3.3V for digital circuits via a low-dropout linear regulator. The second output, +5V / 2A, is an isolated power supply, generating +5V_ISO and ±15V via an isolated DC / DC converter to provide bias voltages for analog circuits and electro-optical conversion. Each output is equipped with transient suppression diodes and LC filters to suppress power supply noise and surges. The power status monitoring circuit monitors each voltage in real time and sends alarm signals when abnormalities occur.

[0041] like Figure 2 As shown, the present invention also provides a method for using an intelligent photoelectric transformer, applied to the aforementioned intelligent photoelectric transformer, comprising: Step 1: The photoelectric acquisition unit senses the primary side current and converts it into an optical signal. The signal conditioning and analog-to-digital conversion module receives the optical signal and converts it into a digital signal sequence. Step 2: The active filtering unit uses a variable step-size normalized least mean square adaptive filter to filter the digital signal sequence. The step-size factor is dynamically adjusted based on the energy of the error signal to output a clean signal. Step 3: The temporal convolutional network state analysis unit extracts features from the clean signal based on a multi-scale dilated temporal convolutional network, and outputs power metering data, power quality monitoring indicators and load trend analysis results in parallel. Step 4: The multi-protocol data communication module encapsulates the results according to the IEC 61850-9-2 protocol and the Modbus TCP / IP protocol respectively, and then sends them to the substation bay layer or station control layer equipment through the fiber optic Ethernet interface.

[0042] Since the photoelectric acquisition unit and signal conditioning and analog-to-digital conversion module have been described and summarized in detail above, step 1 will not be elaborated on here.

[0043] In step 2, the active filtering unit uses a variable step-size normalized least mean square adaptive filter to filter the digital signal sequence. The step-size factor is dynamically adjusted based on the energy of the error signal to output a clean signal. Specifically: Step 2 is executed by the active filtering unit within the data processing and core control module, and its input is the discrete digital signal sequence output from step 1. The output is a clean signal after adaptive filtering. This unit employs an improved variable-step-size normalized least mean square algorithm, which adaptively adjusts the step-size factor according to the characteristics of the power grid signal. While rapidly tracking changes in the power grid, it maintains a low steady-state error, effectively filtering out broadband noise and interference introduced by the photoelectric conversion process and the on-site electromagnetic environment, providing high-quality time-series data for subsequent intelligent analysis. The specific structure, algorithm principle, and parameter settings of this filtering unit are described in detail below.

[0044] 1. Adaptive Filter Structure Design The core of the active filter unit is a unit with a length of A transverse finite impulse response (FIR) digital filter is constructed from a series of delay units, multipliers, and accumulators. In a digital signal processor, this filter is implemented in software. It is defined at discrete time intervals. The filter weight coefficient vector is: ; in, For the first A tap at any time The weighting coefficients. The input signal vector consists of the current and past... Composed of 1 sampling point: ; The value of needs to balance filtering performance and computational complexity, based on the main harmonic components in the power grid (usually analyzed up to the 50th order) and the sampling frequency. ,set up The corresponding time domain length is It is sufficient to cover the signal changes within one fundamental frequency cycle (20ms) and ensure sufficient frequency resolution.

[0045] 2. Generation of the desired response signal Adaptive filtering requires the desired response signal As a learning objective, this invention utilizes phase-locked loop (PLL) technology to track the power grid frequency in real time and generate a sinusoidal reference signal synchronized with the power grid fundamental frequency as the desired response in order to extract a clean power frequency fundamental wave from a noisy signal. The specific implementation is as follows: The digital signal processor core in the data processing and core control module runs a digital phase-locked loop algorithm to... As input, estimate the current power grid frequency in real time. and phase The digital phase-locked loop (PLL) employs a second-order loop filter, enabling it to quickly lock onto frequency changes. Based on the frequency and phase of the PLL output, the desired response is calculated in real time. ; in The sampling interval is denoted by . The desired signal represents a pure fundamental sine wave, serving as the target for the filter approximation. When When harmonics and noise are present, the filter will adjust the weighting coefficients to make the output... Approaching This allows the fundamental component to be extracted, while the error signal is also extracted. This reflects the sum of harmonics and noise.

[0046] 3. Output and Error Calculation At each sampling time The filter first determines the weight vector based on the current weight vector. and input vector Calculation output : ; Then calculate the error signal. : ; error This reflects the deviation between the current filter output and the desired fundamental frequency, and it will be used for subsequent weight coefficient updates.

[0047] 4. Variable Step Size Normalized Least Mean Square Algorithm To overcome the limitation of traditional least mean square (LMS) algorithms where a fixed step size factor cannot simultaneously balance convergence speed and steady-state accuracy, this invention introduces a variable step size mechanism based on error signal energy, combined with normalization processing, to form an improved variable step size normalized least mean square algorithm. The core idea of ​​this algorithm is to adaptively adjust the step size factor according to the instantaneous energy of the error signal, enabling the filter to converge quickly with large step sizes when the error is large, and to finely adjust with small step sizes when the error is small, thereby improving overall performance.

[0048] First, calculate the energy estimate of the error signal. The exponentially weighted moving average method is used: ; in Forgetting factor, range of values The present invention takes This allows energy estimation to smoothly track changes in error and avoid the impact of instantaneous fluctuations.

[0049] Secondly, the current step size factor is dynamically calculated based on the error energy. : ; In the formula The maximum allowable step size determines the maximum convergence speed of the algorithm. This invention takes... ; To adjust the parameters and control the sensitivity of the step size to changes in error energy, take... This functional relationship ensures that when the error energy... When it is large, near This achieves rapid convergence; when the error energy approaches zero, It also approaches zero, effectively reducing steady-state imbalance.

[0050] To prevent drastic changes in the amplitude of the input signal from causing algorithm instability, a normalization factor is introduced, which is the input signal vector. instantaneous energy The final formula for updating the weight coefficients is: ; in It is a very small positive number, used to avoid the denominator being zero. This invention takes... This normalization process makes the step size inversely proportional to the energy of the input signal, ensuring that the algorithm is robust to changes in the amplitude of the input signal.

[0051] 5. Parameter initialization and iteration process At startup, the weight vector Initialize to a zero vector. In each sampling period, the digital signal processor performs the following operations sequentially: Read new sample value Update the input vector ; Based on the output of the phase-locked loop and Calculate the expected response ; Calculate filter output ; Calculation error ; Update error energy estimation ; Calculate variable step size ; Update weight vector .

[0052] After several iterations, the filter converges, and the output... Approaching That is, extracting the fundamental frequency component. However, due to In reality, it's the result of filtering the input signal. It retains harmonic components related to the fundamental frequency (because harmonics are orthogonal to or correlated with the fundamental frequency, the adaptive filter reflects some of them in the error, but doesn't completely eliminate them). In fact, this adaptive filter can be viewed as an adaptive notch filter, and its output... It is the signal after filtering out broadband noise, containing the fundamental frequency and all harmonic components, but with noise effectively suppressed. Therefore, This is called the pure signal, and it serves as the input to the state analysis unit of the subsequent temporal convolutional network.

[0053] The clean signal sequence obtained after processing by the active filter unit The data is continuously stored in another buffer in the dual-port random access memory, which is then used by the temporal convolutional network state analysis unit in step 3 according to a time window. Read. It preserves the time-domain waveform characteristics of the original signal and significantly reduces the noise level.

[0054] In step 3, the temporal convolutional network state analysis unit extracts features from the clean signal based on a multi-scale dilated temporal convolutional network, and outputs in parallel power metering data, power quality monitoring indicators, and load trend analysis results, specifically: This step is performed by the temporal convolutional network state analysis unit within the data processing and core control module, and its input is the clean signal sequence output from step 2. The output includes electricity metering data, power quality monitoring indicators, and load trend analysis results. This unit employs an improved Multi-Scale Dilated Temporal Convolutional Network (MSD-TCN), capable of simultaneously capturing short-term transient features and long-term periodic dependencies in power grid signals, achieving high-precision multi-task parallel output. The following details the network's structural design, operating principles, training process, and output generation method: 1. Overall architecture of multi-scale dilated temporal convolutional networks While traditional temporal convolutional networks can achieve a large receptive field by stacking dilated convolutional layers, their single-scale convolutional kernels struggle to simultaneously capture features at different time scales in power grid signals. For example, the feature scales of harmonic transients (milliseconds) and load trends (minutes) differ significantly. To address this issue, this invention proposes a multi-scale dilated temporal convolutional network, consisting of an input layer, a multi-scale feature extraction module, an attention feature enhancement module, and a task-specific output head.

[0055] The input layer receives the time-series data after filtering in step 2 and extracts it according to a fixed time window. Considering the periodic characteristics of the power grid signal and the requirements of subsequent multi-task analysis, the time window length is set. Each sampling point corresponds to The window length is one second, or five power frequency cycles. This window length ensures sufficient frequency resolution (256 points sampled within the fundamental cycle, allowing analysis up to the 50th harmonic within five cycles) while also capturing short-term trends in load changes. The input tensor is denoted as... ,in 1 represents the number of time steps, and 1 represents the number of characteristic channels (i.e., a single current or voltage signal).

[0056] 2. Design of the multi-scale feature extraction module The multi-scale feature extraction module contains three parallel branches, designed as follows: Branch A (short-term characteristic branch): uses a small inflation factor Three stacked dilated convolutions, kernel size The receptive field of this branch is calculated as follows: Each time step corresponds to a time span of approximately 1.17 ms, and is mainly used to capture short-term disturbance characteristics such as harmonic transients and voltage swells and dips.

[0057] Branch B (Medium-period characteristic branch): Employs a medium expansion factor Three stacked dilated convolutions, kernel size Feel the wild Each time step corresponds to approximately 4.45ms and is primarily used to capture waveform distortion characteristics within the power frequency cycle.

[0058] Branch C (long-period characteristic branch): employs a large expansion factor Three stacked dilated convolutions, kernel size Feel the wild Each time step corresponds to approximately 17.6 ms, which is close to a complete power frequency cycle. It is mainly used to capture long-term trend characteristics such as load fluctuations.

[0059] Each branch's convolutional layers employ causal convolution to ensure temporal directionality, meaning the output at the current time step depends only on the input from the current and past time steps, and not on future information. Causal convolution is implemented by zero-padding the left side of the input sequence before the convolution operation, with a padding length of [value missing]. .

[0060] Each branch employs a residual connection structure to mitigate the gradient vanishing problem in deep networks. Taking branch A as an example, the computation process of its residual block is as follows: ; in For the first The output of the layer, Indicates the first Combination of convolution, normalization, and activation operations in layers To correct the linear unit activation function, a batch normalization layer is added after each convolutional layer to stabilize the training process.

[0061] The outputs of the three branches are concatenated along the channel dimension to obtain multi-scale fused features. ,in The number of output channels for each branch is taken in this embodiment. This splicing operation achieves complementary fusion of features at different time scales: combining the fine features of short-term branches with the macroscopic features of long-term branches, providing rich feature representations for subsequent attention enhancement and task output.

[0062] 3. Design of the attention feature enhancement module To further enhance the representation capability of key temporal features, this invention introduces an efficient convolutional block attention module (ECBAM) after feature fusion. This module consists of a channel attention submodule and a spatial attention submodule connected in series, which can adaptively recalibrate the channel weights and temporal position weights of the features.

[0063] The channel attention submodule is used to model the dependencies between different feature channels. Input feature map First, global average pooling and global max pooling are performed respectively to obtain two channel descriptors. and These two descriptors are each fed into a shared two-layer fully connected network, which consists of a dimensionality reduction layer and a dimensionality increase layer, with a dimensionality reduction ratio of [value missing]. After summing the outputs of the two branches element by element, the channel attention weights are obtained using the Sigmoid function. : ; in This is the Sigmoid function. Compared with the original feature map Multiplying each channel sequentially yields a channel-weighted feature map. .

[0064] The spatial attention submodule is used to model the importance of different temporal locations. Performing average pooling and max pooling along the channel dimension respectively yields two two-dimensional spatial descriptors. and The two descriptors are concatenated along the channel dimension and then fed into a convolutional kernel of size [size missing]. A one-dimensional convolutional layer generates spatial attention weights using the Sigmoid function. : ; Will and Multiplying position by position yields the final enhanced feature map. .

[0065] By cascading channel attention and spatial attention, the network can automatically focus on the most critical feature channels and time-series locations for power metering, harmonic analysis, and load forecasting, significantly improving the model's representation ability and task performance.

[0066] 4. Design of a dedicated output head for the task Feature map after attention enhancement First, a global average pooling layer is used to compress the temporal dimension into a fixed-length feature vector. This vector contains global feature information for the entire time window. Subsequently, The data is fed into three parallel, fully connected output heads, each corresponding to one of the three core tasks of this invention.

[0067] Energy metering output head: This output head consists of two fully connected networks, with a hidden layer dimension of [missing information]. The output layer has 4 dimensions, corresponding to the fundamental voltage amplitude. fundamental current amplitude Phase difference between voltage and current and fundamental frequency Based on these fundamental quantities, active power can be further calculated. and reactive power The output header uses mean squared error as the loss function.

[0068] Power quality monitoring output head: This output head is used to output the total harmonic distortion rate. and the content of each harmonic. Its network structure is also a two-layer fully connected network, with the output layer dimension being... ,in To analyze the highest harmonic order, the first output is... The following Each output represents the amplitude of the harmonic voltage. The formula for calculating the total harmonic distortion rate is: ; in This is the effective value of the fundamental voltage. For the first The effective value of the second harmonic voltage. This output also uses the mean square error loss function.

[0069] Load Trend Analysis Output Header: This output head is used to predict the effective value of the load current in the next time window, employing a sequence-to-sequence prediction mode. Unlike the two output heads mentioned above, this output head uses the attention-enhanced feature map before global average pooling. Take the last one out of the middle The features of each time step (corresponding to 10ms, or half a power frequency cycle) are flattened and fed into a two-layer fully connected network, with the output layer dimension being [dimensional value missing]. That is, the predicted load current values ​​of the next 128 sampling points. For ease of engineering application, the root mean square value of these 128 points can also be used as the prediction of the effective value of the load current in the next time window. The output head uses mean absolute error as the loss function.

[0070] The three output heads operate in parallel without interfering with each other, achieving true multi-task learning. During training, the total loss function is a weighted sum of the loss functions of the three tasks: ; Among them weighting factors Adjustments were made based on the importance and dimensional differences of each task. This invention adopts... ,also, Losses due to electricity metering tasks Losses due to power quality monitoring tasks This is for load trend analysis task loss.

[0071] 5. Model training process and dataset construction To achieve effective training of the aforementioned multi-task temporal convolutional networks, a large-scale, high-quality labeled dataset is required. The dataset is constructed as follows: Data Source: A combination of historical waveform data from actual substations and simulation-generated data was used. The actual data came from three years of historical operation records of a 10kV distribution network in a certain region, including waveform data of normal operating conditions and various disturbance events (voltage swell, sag, interruption, harmonic distortion, flicker, etc.), with a sampling frequency of 12.8kHz. Simulation data was generated according to IEEE Std. 1159-2019 and IEEE Std. 519-2014 standards using MATLAB / Simulink, generating 20 types of power quality disturbance signals and superimposing Gaussian white noise with different signal-to-noise ratios to enhance the robustness of the model.

[0072] Data labeling: For electricity metering tasks, the fundamental amplitude, phase, and frequency of each time window are calculated using standard Fourier analysis as labels; for power quality monitoring tasks, the amplitude of each harmonic and the total harmonic distortion rate are calculated as labels; for load trend analysis tasks, the actual sampled value of the next time window is used as a label.

[0073] Data preprocessing: All signals first undergo adaptive filtering in step 2 to obtain clean signals. Then follow A sliding window is used to capture the time window, with a step size of [missing information]. This process was used to augment the data. The final result was a training set of 200,000 samples, a validation set of 40,000 samples, and a test set of 40,000 samples.

[0074] The network was trained using the Adam optimizer, with an initial learning rate set to... The batch size was 64, and the training epochs were 100. A cosine annealing learning rate strategy was used, with the learning rate decreasing to 0.5 times its original value every 20 epochs. An early stopping mechanism was also introduced, terminating training when the validation set loss did not decrease for 10 consecutive epochs to prevent overfitting. Training was performed on an NVIDIA A100 GPU and convergence took approximately 4 hours.

[0075] 6. Model Inference Process and Output Generation In actual operation, the trained MSD-TCN model is embedded and deployed in the digital signal processor of the data processing and core control module. For each sampling period, the output of step 2... It is continuously stored in the circular buffer. When the buffer is full... An inference calculation is triggered every time a sampling point is reached.

[0076] Input tensor The data is fed into the MSD-TCN network, where it undergoes multi-scale feature extraction, attention enhancement, and global average pooling. The three output heads then calculate the electricity metering data in parallel. Power quality monitoring indicators and load trend analysis results This data is packaged into structured data frames, with a current timestamp attached, and pushed to the multi-protocol data communication module for encapsulation and transmission.

[0077] To ensure real-time performance, inference computation needs to be performed within... Completed within [timeframe]. Through model quantization (converting 32-bit floating-point parameters to 16-bit fixed-point numbers) and computation graph optimization, the actual inference time was controlled within [timeframe]. Within this range, it meets the requirements for real-time processing.

[0078] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0079] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. An intelligent photoelectric transformer, characterized in that, include: The system comprises a photoelectric acquisition unit, a signal conditioning and analog-to-digital conversion module, a data processing and core control module, a multi-protocol data communication module, and a power management module. The photoelectric acquisition unit is connected to the signal conditioning and analog-to-digital conversion module, the signal conditioning and analog-to-digital conversion module is connected to the data processing and core control module, and the data processing and core control module is connected to the multi-protocol data communication module. The photoelectric acquisition unit is used to sense the current in the primary conductive path and convert it into an optical signal output; The signal conditioning and analog-to-digital conversion module is used to receive optical signals and convert them into discrete digital signal sequences; The data processing and core control module includes an active filtering unit and a temporal convolutional network state analysis unit. The active filtering unit uses a variable step-size normalized least mean square adaptive filter to filter discrete digital signal sequences. The variable step-size normalized least mean square adaptive filter dynamically adjusts the step-size factor based on the energy of the error signal to output a clean signal, where the error signal is the difference between the expected response and the actual output of the filter. The temporal convolutional network state analysis unit extracts features from the clean signal based on a multi-scale dilated temporal convolutional network. The multi-scale dilated temporal convolutional network includes at least three parallel branches, each branch using stacked dilated causal convolutional layers with different dilation factor combinations. The output features of each branch are concatenated and processed by an attention enhancement module, and then output through three parallel fully connected output heads to output power metering data, power quality monitoring indicators, and load trend analysis results, respectively. The multi-protocol data communication module is used to encapsulate power metering data, power quality monitoring indicators and load trend analysis results according to IEC 61850-9-2 protocol and Modbus TCP / IP protocol respectively, and then send them to the substation bay layer or station control layer equipment through the fiber optic Ethernet interface. The power management module is used to provide operating voltage for each module.

2. The intelligent photoelectric transformer according to claim 1, characterized in that, The photoelectric acquisition unit includes: a Rogowski coil, an integrator, an electro-optical conversion circuit, and an optical fiber. The Rogowski coil is located on the primary high-voltage bus and connected to the integrator. The integrator is connected to the electro-optical conversion circuit. The electro-optical conversion circuit is connected to the signal conditioning and analog-to-digital conversion module via an optical fiber. The Rogowski coil is used to sense the current being measured and generate an induced electromotive force. The integrator is used to integrate the induced electromotive force to obtain a voltage signal; The electro-optic conversion circuit is used to linearly modulate a voltage signal into an optical signal; The optical fiber is used to transmit optical signals to the signal conditioning and analog-to-digital conversion module.

3. The intelligent photoelectric transformer according to claim 2, characterized in that, The signal conditioning and analog-to-digital conversion module includes: a photodetector, a transimpedance amplifier, an anti-aliasing filter, an analog-to-digital converter, and a digital interface circuit. The optical fiber is connected to the photodetector, the photodetector is connected to the transimpedance amplifier, the transimpedance amplifier is connected to the anti-aliasing filter, the anti-aliasing filter is connected to the analog-to-digital converter, the analog-to-digital converter is connected to the digital interface circuit, and the digital interface circuit is connected to the data processing and core control module. The photodetector is used to convert the received optical signal into photocurrent; The transimpedance amplifier is used to convert photocurrent into a voltage signal; The anti-aliasing filter is a second-order active low-pass filter; The analog-to-digital converter is used to sample and quantize the input voltage and output a 16-bit binary two's complement digital code value. The digital interface circuit is used to transmit 16-bit binary two's complement digital code values ​​to the data processing and core control module in real time via a serial peripheral interface bus.

4. The intelligent photoelectric transformer according to claim 1, characterized in that, The electricity metering data includes the fundamental voltage amplitude, fundamental current amplitude, phase difference, fundamental frequency, and active power calculated accordingly.

5. The intelligent photoelectric transformer according to claim 1, characterized in that, The power quality monitoring indicators include total harmonic distortion rate and voltage amplitude of each harmonic.

6. The intelligent photoelectric transformer according to claim 1, characterized in that, The load trend analysis results include predicted load current values ​​for future sampling points.

7. A method of using an intelligent photoelectric transformer, applied to the intelligent photoelectric transformer according to any one of claims 1-6, characterized in that, include: Step 1: The photoelectric acquisition unit senses the primary side current and converts it into an optical signal. The signal conditioning and analog-to-digital conversion module receives the optical signal and converts it into a digital signal sequence. Step 2: The active filtering unit uses a variable step-size normalized least mean square adaptive filter to filter the digital signal sequence. The step-size factor is dynamically adjusted based on the energy of the error signal to output a clean signal. Step 3: The temporal convolutional network state analysis unit extracts features from the clean signal based on a multi-scale dilated temporal convolutional network, and outputs power metering data, power quality monitoring indicators and load trend analysis results in parallel. Step 4: The multi-protocol data communication module encapsulates the results according to the IEC 61850-9-2 protocol and the Modbus TCP / IP protocol respectively, and then sends them to the substation bay layer or station control layer equipment through the fiber optic Ethernet interface.

8. The method of using the intelligent photoelectric transformer according to claim 7, characterized in that, In step 2, the active filtering unit uses a variable step-size normalized least mean square adaptive filter to filter the digital signal sequence. The step-size factor is dynamically adjusted based on the energy of the error signal to output a clean signal. Specifically: Construct a transverse finite impulse response filter and initialize the weight vector to zero; The desired response is generated by estimating the grid frequency and phase in real time using a digital phase-locked loop. Calculate the output and error of the transverse finite impulse response filter; Calculate the energy estimate of the error signal; The step size factor is dynamically calculated based on the energy estimation of the error signal; The weight vector is updated using the normalized least mean square algorithm; Repeat the above process until convergence is achieved, resulting in a filtered, clean signal.

9. The method of using the intelligent photoelectric transformer according to claim 8, characterized in that, In step 3, the temporal convolutional network state analysis unit extracts features from the clean signal based on a multi-scale dilated temporal convolutional network, and outputs in parallel power metering data, power quality monitoring indicators, and load trend analysis results, specifically: Extract the input tensor from the clean signal according to the time window; Dilated causal convolution is performed on the three parallel branches of the multi-scale feature extraction module to obtain the output features of the three branches, and then the multi-scale fused features are concatenated in the channel dimension. Enhanced features are obtained by performing channel attention and spatial attention enhancement on multi-scale fusion features using an attention feature enhancement module. Global average pooling is used to obtain the feature vector of the enhanced features; The feature vectors are fed into three parallel fully connected output heads to obtain power metering data, power quality monitoring indicators, and load trend analysis results.

10. The method of using the intelligent photoelectric transformer according to claim 9, characterized in that, In step 4, the multi-protocol data communication module encapsulates the results according to the IEC 61850-9-2 protocol and the Modbus TCP / IP protocol respectively, and then sends them to the substation bay layer or station control layer equipment through the fiber optic Ethernet interface. Specifically: The data set of electricity metering data, power quality monitoring indicators and load trend analysis results is constructed according to the IEC 61850-9-2 protocol, encapsulated in Ethernet frames after ASN.1 encoding, and sent to the process layer network through the first optical module in a periodic or event-triggered manner. According to the predefined Modbus register address mapping table, the data is stored in the corresponding register. The second optical module responds to the Modbus TCP / IP read request from the station control layer network, encapsulates the data, and sends it.

Images