Demand response load side control device and method based on neural network

Abstract

The invention discloses a demand response load side control device and method based on a neural network, and the device carries out the multi-scale decomposition and feature analysis of load power data based on a time sequence wavelet LSTM neural network model with an offset, and dynamically distributes demand response power through identifying different load types and the response capability thereof, thereby achieving the control of a demand response load side. And multi-building demand response collaborative optimization is realized based on communication between devices. According to the invention, the problem of insufficient traditional demand response can be solved, and an efficient and reliable load side control device and method are provided for demand response in an intelligent power grid.

Description

Technical Field

[0001] The present invention belongs to the technical field of smart grids, and in particular relates to a demand response load side control device and method based on a neural network. Background Art

[0002] With the increasing energy shortage and environmental problems, the rapid growth of electricity demand has brought tremendous pressure to the operation of the power grid. Traditional power devices are based on the core operation mode of "generation follows load". Due to the increasing intermittency and volatility of renewable energy, the power grid faces severe challenges in balancing supply and demand. In order to improve the stability and economy of power grid operation, "demand response" as an effective load-side regulation strategy has gradually become a research hotspot in the field of smart grids. Demand response encourages users to adjust their electricity consumption behavior, shaving peaks and filling valleys during peak electricity consumption periods, reducing the cost of power grid operation while improving energy efficiency.

[0003] However, the current demand response technology has the following problems: first, the real-time and accuracy of load regulation are poor, making it difficult to efficiently respond to dynamic grid demand; second, there is a lack of intelligent optimization methods for different buildings or load characteristics, which limits the effect of demand response; third, traditional control devices lack flexibility in multi-building collaborative decision-making and cross-regional load balancing, and cannot fully tap the potential of demand response. Therefore, how to design an intelligent and efficient load-side control device to dynamically optimize power distribution and coordinate decisions between multiple devices has become an important direction for demand response. Summary of the invention

[0004] In view of the above problems, the present invention provides a neural network-based demand response load-side control device and method, aiming to solve the problems of poor real-time performance, insufficient accuracy and weak multi-building collaborative optimization capability of load regulation in existing demand response technologies.

[0005] According to an embodiment of the present disclosure, a neural network-based demand response load-side control device is provided. The device performs multi-scale decomposition and feature analysis on load power data based on a time-series wavelet LSTM neural network model with an offset, dynamically allocates demand response power by identifying different load types and their response capabilities, and realizes collaborative optimization of multi-building demand response based on communication between devices.

[0006] A further technical solution of the present invention is as follows: The device includes a hardware unit and a software unit. The hardware unit includes a data collection module, an inter-device communication module, a control signal module, and an FPGA module. The software unit includes a load identification module, an available load estimation module, a single control device power distribution module, a multi-control device collaborative power distribution module, and a demand response effect evaluation module. Among them, the control signal module is used to control the load to receive and execute the demand response. The software unit is deployed on the FPGA module, and the FPGA module is used to provide logical operations, temporary data storage, and data interfaces.

[0007] A further technical solution of the present invention is as follows: The data collection module supports manual input of the rated power of controllable loads and real-time collection of voltage, current, and power data. The sampling frequency is not less than once every ten seconds, and the storage capacity is not less than 500MB.

[0008] A further technical solution of the present invention is as follows: The inter-device communication module supports Bluetooth, 5G, and Ethernet communications, automatically selects the communication method with the minimum delay, and the maximum communication bandwidth is 1Gbps.

[0009] A further technical solution of the present invention is as follows: The FPGA module contains no less than 500,000 logic units, the on-chip memory capacity is not less than 2MB, and the clock frequency is not less than 100MHz.

[0010] A further technical solution of the present invention is as follows: The load identification module uses Fourier transform to decompose the heating ventilation and air conditioning (HVAC) load and lighting load in the building electricity consumption data, and combines neural network and time series analysis algorithms to obtain the typical power curves of different load types operating. The available load estimation module evaluates the available power capacity based on the real-time power, rated power, and typical power curves, and preferentially adjusts the load types with less impact on user comfort.

[0011] A further technical solution of the present invention is as follows: The single control device power distribution module performs power distribution on controllable loads based on a time series wavelet long short-term memory (LSTM) neural network model with an offset. Based on the external demand response power change requirements, it distributes the demand response power to each controllable load in the building. The distribution result is the ratio of the response power to the rated power of the load.

[0012] A further technical solution of the present invention is as follows: The multi-control device collaborative power distribution module uses a distributed algorithm to achieve global optimization across buildings, dynamically adjusts the response weights of each device, and preferentially adjusts the devices with more idle capacity and lower response costs.

[0013] A further technical solution of the present invention is as follows: The demand response effect evaluation module evaluates the demand response effect through the response time and steady-state error. The response time is the time required for the power to stabilize after the instruction is issued, and the steady-state error is the deviation between the actual power and the target power.

[0014] According to another embodiment of the present disclosure, a demand response load - side control method based on a neural network is provided. The method includes performing multi - scale decomposition and feature analysis on load power data based on a time - series wavelet LSTM neural network model with an offset, dynamically allocating demand response power by identifying different load types and their response performances, and achieving collaborative optimization of multi - building demand response based on device - to - device communication.

[0015] A further technical solution of the present invention is that, based on the external requirement of demand response power change, the method allocates demand response power to each controllable load in a building, and the allocation result is the ratio of the response power to the rated power of the load.

[0016] A further technical solution of the present invention is that the method evaluates the demand response effect through response time and steady - state error. The response time is the time required for the power to stabilize after the command is issued, and the steady - state error is the deviation between the actual power and the target power.

[0017] A further technical solution of the present invention is that the method uses Fourier transform to decompose the heating, ventilation, and air - conditioning (HVAC) load and lighting load in the building electricity consumption data, and combines a neural network and a time - series analysis algorithm to obtain the power curves of different load types operating;

[0018] Based on the real - time power, rated power, and typical power curves, the available power capacity is evaluated, and the load types with less impact on user comfort are preferentially adjusted.

[0019] A demand response load - side control device and method based on a neural network provided by an embodiment of the present disclosure. The device performs multi - scale decomposition and feature analysis on load power data based on a time - series wavelet LSTM neural network model with an offset, dynamically allocates demand response power by identifying different load types and their response capabilities, and achieves collaborative optimization of multi - building demand response based on device - to - device communication. The beneficial effects are as follows:

[0020] Based on the time - series wavelet LSTM neural network technology with an offset, through the accurate identification of load response characteristics, the dynamic optimization of load - side demand response is realized, improving the real - time performance and accuracy of the response.

[0021] Supports communication and data sharing between multiple control devices. Combining with a distributed optimization algorithm, the overall adjustment efficiency is improved in the multi - building demand response scenario.

[0022] The data acquisition, processing, and storage are completed through an FPGA module, supporting real - time calculation of demand response effect indicators, including response time and steady - state error.

[0023] Supports multiple communication methods such as Bluetooth, 5G, and Ethernet, and can dynamically select the optimal communication method according to the network environment to ensure the efficiency and reliability of data transmission.

[0024] In summary, the present invention can solve the deficiencies of traditional demand response technologies and provide an efficient and reliable load-side control device and method for demand response in smart grids.

[0025] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings herein are incorporated into the specification and constitute a part of this specification, showing embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention.

[0027] Figure 1 is a schematic diagram of the hardware unit structure in the demand response load-side control device based on neural network according to an embodiment of the present invention;

[0028] Figure 2 is a schematic diagram of the software unit structure in the demand response load-side control device based on neural network according to an embodiment of the present invention;

[0029] Figure 3 is a schematic diagram of the structure of the time-series wavelet LSTM neural network model with an offset according to an embodiment of the present invention;

[0030] Figure 4 is a schematic diagram of the result of the load identification module according to an embodiment of the present invention;

[0031] Figure 5 is a schematic diagram of two physically connected control devices according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present invention, rather than limiting the present invention. Additionally, it should be noted that for the sake of convenience of description, only parts related to the present invention rather than all structures are shown in the drawings.

[0033] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts depict the steps as sequential processes, many of the steps can be implemented in parallel, concurrently, or simultaneously. In addition, the order of the steps can be rearranged. The process can be terminated when its operations are completed, but it can also have additional steps not included in the drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0034] Aiming at the problems existing in the existing demand response technologies, such as insufficient load regulation accuracy, poor real-time performance, and weak multi-building collaborative optimization ability, the purpose of the present invention is to provide a demand response load-side control device and method based on a time-series wavelet LSTM neural network with an offset. This device can achieve real-time optimal control on the load side, dynamically allocate demand response power, and at the same time support collaborative optimization among multiple control devices, comprehensively improving the demand response effect and the grid operation efficiency.

[0035] In the process of load-side control by traditional demand response technologies, power distribution to loads usually depends on fixed rules or simple models, which is difficult to adapt to the complex and changeable grid operation environment, and lacks accurate identification and control of the response characteristics of different types of loads. In addition, in the scenario of multi-building joint demand response, existing devices lack efficient communication and collaborative decision-making capabilities, resulting in low response efficiency and difficulty in meeting the requirements of modern smart grids. To address these problems, the present invention proposes a demand response control device based on a time-series wavelet LSTM neural network with an offset, and solves the technical problems of real-time optimal control on the load side of demand response and collaborative optimization of multiple devices by introducing the combined technology of wavelet transform and neural network.

[0036] Due to its powerful time-frequency analysis ability and non-linear modeling ability, the time-series wavelet LSTM neural network with an offset has received extensive attention in the field of load prediction and control. By combining the time-series wavelet LSTM neural network with an offset and demand response technologies, the real-time performance and accuracy of load response can be effectively improved, and at the same time, intelligent identification and optimal control of different load types can be achieved. Based on this, the present invention proposes a demand response load-side control device based on a time-series wavelet LSTM neural network with an offset, solves the deficiencies of existing demand response technologies, and comprehensively improves the grid operation efficiency and load regulation ability.

[0037] An embodiment is a demand response load-side control device and method based on a time-series wavelet LSTM neural network with an offset, especially relating to a control device for building loads participating in demand response, aiming to solve the problems existing in the existing demand response technologies, such as poor real-time performance of load regulation, insufficient accuracy, and weak multi-building collaborative optimization ability. The present invention has the advantages of strong real-time performance, precise optimization, and high collaborative ability, and is applicable to the demand response scenario of smart grids.

[0038] The device of the present invention performs multi-scale decomposition and feature analysis on load power data based on a time-series wavelet LSTM neural network model, dynamically allocates demand response power by identifying different load types and their response capabilities, and realizes collaborative optimization of multi-building demand response based on communication between devices.

[0039] Specifically, the device allocates demand response power for each controllable load in the building according to the requirements of the whole building's demand response power change proposed by the outside world. The allocation result is given in the form of the ratio of the response power to the rated power of the load. The response performance of each load is optimized through a time series wavelet LSTM neural network model with an offset, and the adjustment strategy of the load is dynamically adjusted according to the output result of the model to achieve the real-time and accuracy of power allocation.

[0040] As Figure 1 , Figure 2 shown, the device includes a hardware unit and a software unit. The hardware unit includes a data collection module, an inter-device communication module, a control signal module, and an FPGA module. The software unit includes a load identification module, an available load estimation module, a single control device power allocation module, a multi-control device collaborative power allocation module, and a demand response effect evaluation module. Among them, the control signal module is used to control the load to receive and execute the demand response. The software unit is deployed on the FPGA module, and the FPGA module is used to provide logical operations, temporary data storage, and data interfaces.

[0041] The data collection module supports manual input of the rated power of the controllable load and real-time collection of voltage, current, and power data. The sampling frequency is not less than once every ten seconds, and the storage capacity is not less than 500MB.

[0042] Specifically, the data collection module provides a function for manually inputting the rated power of the controllable load; can sample, store, and upload the real-time power; and includes a non-volatile memory with a storage capacity of not less than 500MB.

[0043] The inter-device communication module supports Bluetooth, 5G, and Ethernet communications, automatically selects the communication method with the minimum delay, and the maximum communication bandwidth is 1Gbps. Ensure the high efficiency of data transmission through various communication methods such as Bluetooth, 5G, and Ethernet.

[0044] Specifically, the devices are connected to each other through three hardware facilities: a Bluetooth module, a 5G module, or a network cable interface. The inter-device communication module of the control device supports Bluetooth protocol, TCP / IP protocol, and 5G communication standard. After startup, the module can automatically judge the best communication method and preferentially select the one with the minimum communication delay for data transmission when multiple communication methods are available.

[0045] The control signal module can directly control the on / off state and brightness adjustment of the lighting load, and the on / off state and temperature adjustment of the HVAC load through hardware facilities. The control signal module can enable the controllable load to receive and execute the demand response strategy, and at the same time, the load status is feedback in real time through the data collection module, and the sampled data is transmitted to the FPGA module.

[0046] The FPGA module contains no less than 500,000 logic units, the on-chip memory capacity is not less than 2MB, and the clock frequency is not less than 100MHz.

[0047] The load identification module uses Fourier transform to decompose the heating ventilation and air conditioning (HVAC) load and lighting load in the building electricity consumption data, and combines neural network and time series analysis algorithms to obtain the typical power curves of different load types in operation; the available load estimation module evaluates the available power capacity based on the real-time power, rated power and typical power curves, and preferentially adjusts the load types with less impact on user comfort.

[0048] Specifically, the load identification module reads the building electricity consumption data and uses Fourier transform to decompose the HVAC load and lighting load; the identification method combines neural network and time series analysis algorithms to obtain the typical power curves of different load types in daily operation, providing data support for subsequent demand response strategies. It can evaluate the available power capacity through real-time power, rated power and load typical power curves; the available load estimation module distributes the maximum possible demand response power to each load on the premise of ensuring the rigid demand of building electricity consumption; the module establishes a load priority ranking model and preferentially adjusts the load types with less impact on user comfort.

[0049] The single control device power distribution module distributes power to controllable loads based on the time series wavelet LSTM neural network model with offset; based on the external requirement of demand response power change, it distributes demand response power to each controllable load in the building, and the distribution result is the ratio of the response power to the rated power of the load.

[0050] Specifically, the single control device power distribution module distributes power to the controllable loads within its jurisdiction based on the time series wavelet LSTM neural network model with offset; the module calculates the ratio of the response power of each load to the rated power by receiving the demand response instruction and combining the output data of the load identification module and the available load estimation module; the distribution strategy uses an optimization algorithm to comprehensively consider the accuracy and response time of power distribution to minimize the target error while ensuring user comfort and electricity demand; the distribution result is directly sent to the relevant load equipment through the control signal module and updated in real time.

[0051] The multi-control device collaborative power distribution module adopts a distributed algorithm to achieve global optimization across buildings, dynamically adjusts the response weights of each device, and preferentially adjusts the devices with more idle capacity and lower response cost.

[0052] Specifically, the multi-control device collaborative power allocation module is used to achieve collaborative optimization across control devices in multi-building demand response scenarios; the module establishes a real-time data sharing and joint decision-making mechanism among multiple devices through the communication module between control devices, and adopts a distributed power allocation algorithm based on the local load status and available capacity provided by each control device to optimize the demand response power allocation on a global scale; the module can dynamically adjust the response weights between control devices, give priority to adjusting devices with more idle capacity and lower response cost, and improve the overall demand response efficiency and reliability.

[0053] The demand response effect evaluation module evaluates the demand response effect through response time and steady-state error. The response time is the time required for the power to stabilize after the command is issued, and the steady-state error is the deviation between the actual power and the target power. The two indicators of response time and steady-state error are used to comprehensively evaluate the demand response effect and improve the real-time and accuracy of the response.

[0054] Specifically, the response time is used to characterize the load response speed, that is, the time required from the issuance of the demand response instruction to the load power reaching the stable target value; the steady-state error is used to evaluate the accuracy between the load response power and the target power, that is, the deviation between the actual response power of the load in a stable state and the expected response power.

[0055] The embodiment controls the controllable loads in the building to participate in demand response accurately and quickly based on the collection and processing of electricity consumption data. The device consists of hardware units and software units. The hardware unit is the basis for data transmission and control signal transmission. Figure 1 The framework of the hardware unit of the device proposed by the present invention is shown. The software unit realizes generating a power regulation signal for each load by responding to the power regulation signal through historical power consumption data, real-time power consumption data and the total demand of the building. Figure 2 The framework of the software unit of the device proposed by the present invention is shown.

[0056] The data collection module of the device provides both the function of manually inputting the rated power of the controllable load and the automatic collection of real-time power data of the load in the building, and stores historical data in non-volatile memory. Among them, the automatic sampling frequency is not less than once every ten seconds to ensure that the load power fluctuation is fully recorded. The sampling power range is 0.1kW to 30kW. The data collection module calculates the real-time power of the load by sampling the instantaneous values ​​of voltage and current, and saves the data to the SD memory card. When the storage space of the SD memory card overflows, the original stored data is sequentially rolled over.

[0057] The maximum communication bandwidth supported by the communication module between control devices is 1 Gbps, which can meet the requirements of high-frequency data exchange among multiple devices; the Bluetooth protocol supported by the device is Bluetooth 5.0, the TCP / IP protocol is IPv6, and the 5G communication standard is NR (New Radio) Release 16; the maximum communication distance of Bluetooth is not less than 30 meters, and the maximum communication distance of 5G communication is not less than 100 meters. The transmission delay of the control signal sent by the control signal module of the device does not exceed 20 ms to ensure the real-time performance of load response. The output voltage range of the control signal is 3.3V to 5V, which is compatible with various load control interfaces. To support the operation of the software unit, the number of logic units (LUTs) included in the FPGA module is not less than 500,000 to support the calculation of complex time-series wavelet LSTM neural network models with offsets; the on-chip memory capacity is not less than 2 MB for real-time storage of intermediate calculation data; the clock frequency is not less than 100 MHz to ensure high computing power; the supported power consumption control range is 5W to 15W to meet the requirements of low-power operation.

[0058] The software unit is deployed on the FPGA module in the hardware unit. In the software unit, the brightness adjustment range for the lighting load is 10% to 100%, and the temperature adjustment range for the HVAC load is 16°C to 30°C, with an adjustment accuracy of ±0.5°C. The load identification module uses Fourier transform to decompose the total building load curve into a combination of sub-curves and finds the sub-curves consistent with the usage characteristics of lighting and HVAC as typical load curves. The available load estimation module determines the scale of real-time available load through the typical load curve, current usage power, and rated power of controllable loads. After receiving the demand response signal, in the mode of manually setting single control devices or coordinating multiple control devices, considering the electricity consumption experience of users in the building, relying on the time-series wavelet LSTM neural network with offsets to complete the optimal power distribution and modulate the control signal to be sent to each load for execution. Finally, the demand response effect evaluation module evaluates the response speed and steady-state error of this response.

[0059] Figure 3 Shows the structure of the time-series wavelet LSTM neural network with offsets. The input of the neural network is a 3*t*n matrix. The first dimension of the matrix represents 3 groups of different time-series data. One group is normal time-series data, and the other 2 groups are different time-series data after time-series offset processing. For example, assume the sequence without time-series offset processing is x 1 ,x 2 ,...,x t , then the other two groups are x 0 ,x 1 ,...,x t-1 and x 2 ,x 3 ,...,x t+1The second dimension represents the most recent t sampling points in the past. The third dimension represents the ratio of the operating power at which each controllable load is adjusted to its rated power. Here, n represents the number of controllable loads. After the input data enters the neural network, it first undergoes discrete wavelet transform to extract the effective features of the first 6 layers, and then is sent to other structures of the neural network for processing by different LSTM neurons. In the present invention, the discrete wavelet transform uses DB4 as the wavelet function. The output of the entire neural network is a 1*n matrix, representing the ratio of the power adjusted for each controllable load at the next moment to its total power.

[0060] Another embodiment is a demand response load - side control method based on a neural network. The method includes performing multi - scale decomposition and feature analysis on load power data based on a time - series wavelet LSTM neural network model with an offset, dynamically allocating demand response power by identifying different load types and their response performances, and realizing collaborative optimization of multi - building demand response based on device - to - device communication.

[0061] Further, based on the external requirements for changes in demand response power, the method allocates demand response power to each controllable load in the building, and the allocation result is the ratio of the response power to the rated power of the load.

[0062] Further, the method evaluates the demand response effect through response time and steady - state error. The response time is the time required for the power to stabilize after the command is issued, and the steady - state error is the deviation between the actual power and the target power.

[0063] Further, the method uses Fourier transform to decompose the heating, ventilation, and air - conditioning (HVAC) load and lighting load in the building's electricity consumption data, combines the neural network with the time - series analysis algorithm to obtain the power curves of different load types operating; based on the real - time power, rated power, and typical power curves, evaluates the available power capacity, and preferentially adjusts the load types with less impact on user comfort.

[0064] For the other specific working processes of the demand response load - side control method based on a neural network, refer to the description of the embodiment of the demand response load - side control device based on a neural network above, and details will not be repeated.

[0065] In a specific embodiment, based on the control device A of the present invention, the smart electricity meter can detect the load usage situation in the building and has collected historical data for the past 30 days. Controllable lamp groups, HVAC systems, and other non - controllable loads. The control device of the present invention automatically identifies the load status and evaluates the adjustable load capacity.

[0066] In this example, during the operation of the control device of the present invention, first, the data collection module is connected to the smart electricity meter to collect the voltage, current, and power data in the building in real time, and perform preprocessing in combination with the stored historical data. The control device A performs multi-dimensional decomposition on the collected load data through Fourier transform and time series analysis techniques to identify three types of loads in the building: controllable lamp groups, HVAC systems, and other uncontrollable loads. For controllable loads, the control device extracts their typical power curves through a time series wavelet LSTM neural network model with an offset, and automatically identifies the operating states of different loads. For example, the lighting load is identified as a controllable load through the characteristics of brightness change and power fluctuation, while the HVAC system is classified as another type of controllable load through its periodic start-stop characteristics. For other uncontrollable loads, such as office equipment and household appliances, they are identified as uncontrollable loads through their power time series characteristics. Figure 4 Shows the results of load identification.

[0067] In another specific embodiment, a control device proposed by the present invention is installed in an office building, and this device is responsible for adjusting the controllable loads in the building to meet the demand response instructions of the power dispatching center. The total load of the building includes two groups of lighting loads, an HVAC system, and office equipment loads. According to the instructions of the dispatching center, the building needs to reduce 5 kW of power within 15 minutes. The control device needs to identify the controllable loads, calculate the response ratio, and achieve the goal through instruction issuance. The specific data is shown in Table 1.

[0068] Table 1 Data table of controllable loads

[0069]

[0070] In this example, after the control device of the present invention is started, it first collects the voltage, current, and power data in the building in real time through the data collection module, and decomposes the total power consumption curve through Fourier transform to find the characteristics corresponding to the lighting load and the HVAC load. Through historical data and electricity consumption characteristic analysis, Lighting Load 1 and Lighting Load 2 are identified as fast controllable loads, the HVAC load is identified as a controllable load with a large capacity but slow response, and office equipment is classified as an uncontrollable load due to its non-adjustable characteristics and does not participate in demand response. According to the evaluation results of the control device, the maximum adjustable power of Lighting Load 1 is 1.0 kW, the maximum adjustable power of Lighting Load 2 is 0.4 kW, the maximum adjustable power of the HVAC system is 1.1 kW, and the total adjustable power is 2.5 kW. Since the total adjustable power is not enough to meet the demand response goal (5 kW), the control device starts a time series wavelet LSTM neural network model with an offset to optimize the calculation of the load response ratio, so as to dynamically adjust the adjustment intensity of each load.

[0071] According to the optimized calculation results, the device allocates the response power shown in Table 2 to each load respectively:

[0072] Table 2 Allocation Response Power Table

[0073] Load type Distributed response power (kW) Proportion of distributed power to rated power (%) Lighting load 1 0.8 26.7％ Lighting load 2 0.4 20.0％ HVAC load 3.8 63.3％

[0074] In this allocation scheme, the lighting load 1 is reduced to operate at 1.7 kW, and the lighting load 2 is reduced to operate at 1.4 kW. The HVAC system reduces the power consumption by 3.8 kW by increasing the set temperature, thus meeting the demand response target of a total power reduction of 5 kW. The control device preferentially allocates the lighting load with less impact on user comfort and takes the HVAC system as the main load reduction object to achieve a larger power adjustment space. After the instruction is issued and the adjustment is completed, the control device starts the demand response effect evaluation module to monitor the response effect in real time. The evaluation results show that the response time from the instruction issuance to the completion of the adjustment of all loads is 1 second, significantly lower than the upper limit of the demand response delay; the finally reduced power is 5.02 kW, and the steady-state error is +0.4%, meeting the error requirements. In addition, the adjustment of the allocation ratio has little impact on the actual electricity consumption experience of users. The adjustment of the lighting load does not cause obvious discomfort, and the increase in the temperature of the HVAC system is still within the acceptable range of users.

[0075] In another embodiment, in an office park containing two buildings, two control devices A and B of the present invention are respectively installed, and each control device is responsible for managing the controllable loads in its respective building. The load conditions and related parameters of the two buildings are shown in Table 3:

[0076] Table 3 Load-Related Parameter Table

[0077]

[0078] The dispatching center issues a demand response instruction, requiring two buildings to jointly reduce the total power by 8 kW within 15 minutes. The control devices need to share data through the communication module, allocate their respective load response ratios through the distributed optimization algorithm, and complete the adjustment tasks. In this example, control device A is responsible for the load management of Building 1, and control device B is responsible for the load management of Building 2. Through the communication module between the control devices, the two control devices share their respective load status data and jointly calculate their respective reduction ratios and response power allocations. After the demand response instruction is issued, the two control devices start the data collection module, respectively collect the real-time load data in their respective buildings, and combine the historical data of the past 30 days to evaluate the load response capacity. The evaluation results show that the total adjustable power of Building 1 is 5.6 kW, the total adjustable power of Building 2 is 5.9 kW, and the total adjustable power of the two buildings is 11.5 kW, which is sufficient to meet the demand of reducing 8 kW. To optimize the reduction allocation, control devices A and B jointly calculate the reduction targets through the distributed optimization algorithm. According to the proportion of the adjustable power of each building, the reduction target of Building 1 is allocated as 3.8 kW, and the reduction target of Building 2 is allocated as 4.2 kW. Subsequently, each control device optimally allocates the load in its own building, and the specific allocation results are shown in Table 4:

[0079] Table 4 Load Allocation Response Power Table

[0080]

[0081] According to the allocation results, the lighting load 1 of Building 1 is reduced to 2.0 kW for operation, the HVAC load 1 is reduced to 4.0 kW for operation, and the HVAC load 2 is reduced to 3.1 kW for operation, with a total reduction power of 3.8 kW. The lighting load 2 of Building 2 is reduced to 2.6 kW for operation, the lighting load 3 is reduced to 5.7 kW for operation, and the HVAC load 3 is reduced to 2.4 kW for operation, with a total reduction power of 4.2 kW. After the reduction instruction is issued, the control device starts the demand response effect evaluation module to monitor the adjustment process in real time. The final evaluation results show that the response time from the instruction issuance to the completion of the adjustment of all loads is 1.2 seconds, which is significantly lower than the 15-second upper limit set for the demand response. The total reduction power of the two buildings is 8.02 kW, and the steady-state error is +0.25%, which is far lower than the allowable error range. In addition, through the distributed optimization algorithm, each control device achieves efficient cooperation without centralized calculation. The allocation results not only meet the demand response objectives but also minimize the impact on user comfort and normal power consumption to the greatest extent.

[0082] In another embodiment, there are two control devices A and B of the present invention. Both are equipped with a Bluetooth module (Bluetooth 5.0), a 5G module (NR Release 16), and a network cable interface (gigabit Ethernet), and provide communication conditions for Bluetooth, 5G, and wired network respectively in a simulated network environment. After the communication methods of the two control devices are physically connected and started, the communication method with the shortest delay is automatically selected. The connection method is as Figure 5 shown. In this example, control device A sets a data packet transmission task of 512 bytes for each communication method, and each communication method is transmitted 100 times to calculate the average delay. The delay measurement method is the time required from control device A to send a data packet to control device B to successfully receive the data packet. The delay result takes the average value of 100 transmissions as the final reference. After device A is started, the communication module between control devices first detects the availability of the three communication methods and measures the delay for each method in turn. The measurement results show that the average delay of the Bluetooth module is 30 ms, the average delay of the 5G module is 15 ms, and the average delay of the network cable interface is the smallest, only 5 ms. According to the actually measured delay data, the communication module between control devices can judge that the delay of the network cable interface is the smallest. When multiple communication methods are available, the network cable interface is automatically selected as the communication method for data transmission. Subsequently, the data packet is transmitted through the network cable interface, and the data is stable during the transmission process and no packet loss occurs, verifying the correctness of the selection result.

[0083] In this article, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a step or method comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such step or method.

[0084] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, which should all be regarded as belonging to the protection scope of the present invention.

Claims

1. A demand response load side control device based on neural network, characterized in that: The device performs multi-scale decomposition and feature analysis on load power data based on a time-series wavelet LSTM neural network model with an offset, dynamically allocates demand response power by identifying different load types and their response capabilities, and realizes collaborative optimization of multi-building demand response based on communication between devices.

2. The neural network-based demand response load side control device according to claim 1, characterized in that: The device includes a hardware unit and a software unit. The hardware unit includes a data collection module, an inter-device communication module, a control signal module and an FPGA module; the software unit includes a load identification module, an available load estimation module, a single control device power allocation module, a multi-control device collaborative power allocation module and a demand response effect evaluation module, wherein the inter-device communication module is used to achieve collaborative optimization of multi-building demand response, the control signal module is used to control the load to receive and execute demand response, and the software unit is deployed in the FPGA module, and the FPGA module is used to provide logical operations, temporary data storage and data interface.

3. The neural network-based demand response load side control device according to claim 2, characterized in that: The data collection module supports manual input of the rated power of the controllable load and collects voltage, current and power data in real time, with a sampling frequency of not less than once every ten seconds and a storage capacity of not less than 500MB.

4. The neural network-based demand response load side control device according to claim 2, characterized in that: The inter-device communication module supports Bluetooth, 5G and Ethernet communications, automatically selects the communication method with the least delay, and the maximum communication bandwidth is 1Gbps.

5. The neural network-based demand response load side control device according to claim 2, characterized in that: The FPGA module contains no less than 500,000 logic units, an on-chip memory capacity of no less than 2MB, and a clock frequency of no less than 100MHz.

6. The neural network-based demand response load side control device according to claim 2, characterized in that: The load identification module uses Fourier transform to decompose the HVAC load and lighting load in the building electricity consumption data, and combines neural network and timing analysis algorithms to obtain typical power curves for different load types. The available load estimation module evaluates the available power capacity based on real-time power, rated power and typical power curves, and gives priority to adjusting load types that have less impact on user comfort.

7. The neural network-based demand response load side control device according to claim 2, characterized in that: The single control device power allocation module allocates power to the controllable load based on a time-series wavelet LSTM neural network model with an offset; based on the external demand response power change requirements, the demand response power is allocated to each controllable load in the building, and the allocation result is the ratio of the response power to the load rated power.

8. The neural network-based demand response load side control device according to claim 2, characterized in that: The multi-control device collaborative power allocation module adopts a distributed algorithm to achieve global optimization across buildings, dynamically adjusts the response weight of each device, and gives priority to adjusting devices with more idle capacity and lower response cost.

9. The neural network-based demand response load side control device according to claim 2, characterized in that: The demand response effect evaluation module evaluates the demand response effect through response time and steady-state error. The response time is the time required for power stabilization from the issuance of the instruction, and the steady-state error is the deviation between the actual power and the target power.

10. A demand response load side control method based on neural network, characterized in that: The method includes multi-scale decomposition and feature analysis of load power data based on a time-series wavelet LSTM neural network model with an offset, dynamically allocating demand response power by identifying different load types and their response performance, and realizing collaborative optimization of multi-building demand response based on communication between devices.

Images