Distributed optical storage micro-grid system optimization method

Abstract

The invention discloses a distributed optical storage micro-grid system optimization method, and the method comprises the steps: applying a CEA module to a decomposition model of non-intrusive load monitoring NILM in the field of distributed optical storage micro-grid systems, and extracting the load of a single household electrical appliance based on a local cross-channel interaction load decomposition method, therefore, prediction characteristics and a scheduling basis are provided for load prediction through the load condition, more accurate prediction and more effective energy scheduling are helped to be realized, the contradiction between the performance of a channel attention mechanism in a non-invasive load decomposition model and the model complexity is solved, the accuracy of load decomposition is remarkably improved, and the load decomposition efficiency is improved. And the load decomposed by the NILM is used as a feature, so that a basis of more dimensions is provided for load prediction of the photovoltaic energy storage system, continuous power supply of important loads is guaranteed, and the energy utilization cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of distributed photovoltaic and energy storage microgrids, and in particular to a distributed photovoltaic and energy storage microgrid system optimization method. Background Art

[0002] The widespread use of primary energy sources such as oil and coal has promoted economic development and facilitated people's daily lives, but has also caused widespread environmental and climate damage. Photovoltaic power generation, as a key clean energy source, is intermittent and random, but through the application of energy storage technology, it can effectively achieve stable output and reduce the impact of power fluctuations, thereby promoting the stable operation and widespread application of photovoltaic storage microgrids, helping to solve the bottleneck of photovoltaic access and reduce equipment requirements, and promoting the green transformation of the energy structure.

[0003] Distributed photovoltaic microgrid systems, as energy systems that integrate solar photovoltaic power generation, energy storage and microgrid technologies, are becoming increasingly important in home and commercial environments.

[0004] According to some information, taking the household distributed photovoltaic storage microgrid system as an example, the main function of the household distributed photovoltaic storage microgrid system is to reduce the cost of energy utilization. The energy storage equipment is charged through photovoltaic power generation and grid charging. The distributed photovoltaic storage microgrid system needs to be charged from the grid due to the uncertainty of photovoltaic power generation, and the volatility of grid prices in different time periods will lead to different charging costs. Therefore, how to dispatch energy has become the focus of the distributed photovoltaic storage microgrid system. Through energy dispatching, not only the charging cost can be reduced, but also the electricity cost can be reduced to ensure the continuous power supply of important loads. What determines the energy dispatch is the load power consumption, so household load forecasting has become a crucial step.

[0005] Chinese patent document CN111416390B discloses a "distributed photovoltaic energy storage microgrid control method". The microgrid includes a distributed photovoltaic array, a DC / DC switching system, an inverter control system, an energy storage system, a load and a distribution network; the microgrid has three working modes: grid-connected operation mode, off-grid operation mode and system recovery grid-connected operation mode. The above technical solution is difficult to accurately extract the load of the household distributed photovoltaic storage microgrid system and accurately dispatch it. Summary of the invention

[0006] The present invention mainly solves the technical problem that it is difficult to accurately extract the load of a household distributed photovoltaic storage microgrid system and accurately schedule it, and provides a distributed photovoltaic storage microgrid system optimization method. In the field of distributed photovoltaic storage microgrid systems, the ECA module is applied to the decomposition model of non-intrusive load monitoring NILM, and the load of a single household appliance is extracted based on the method of local cross-channel interactive load decomposition, so as to provide prediction characteristics and scheduling basis for load forecasting through the load situation, and help achieve more accurate prediction and more effective energy scheduling. It not only solves the contradiction between the performance of the channel attention mechanism and the complexity of the model in the non-intrusive load decomposition model, but also significantly improves the accuracy of load decomposition, and uses the load decomposed by NILM as a feature, which provides a multi-dimensional basis for load forecasting of photovoltaic energy storage systems, ensures the continuous power supply of important loads, and reduces energy utilization costs.

[0007] The above technical problem of the present invention is mainly solved by the following technical solution: The present invention comprises the following steps: S1 constructing a neural network model of non-intrusive load decomposition based on the ECA module; S2 uses a neural network model to decompose the household distributed photovoltaic storage microgrid system to obtain the power of a single appliance; S3 obtains the load forecast of a single appliance by taking the decomposed power situation of the single appliance as the feature; S4 performs charging and discharging scheduling according to the predicted load conditions.

[0008] In the field of distributed photovoltaic and energy storage microgrid systems, the ECA module is applied to the decomposition model of non-intrusive load monitoring NILM. The load of a single household appliance is extracted based on the method of local cross-channel interactive load decomposition, so as to provide prediction characteristics and scheduling basis for load forecasting through the load situation, helping to achieve more accurate prediction and more effective energy scheduling. It not only solves the contradiction between the performance of the channel attention mechanism and the complexity of the model in the non-intrusive load decomposition model, but also significantly improves the accuracy of load decomposition. The load decomposed by NILM is used as a feature, which provides a more dimensional basis for load forecasting of photovoltaic energy storage systems, ensures the continuous power supply of important loads, and reduces energy utilization costs.

[0009] Preferably, the neural network model in step S1 includes an encoder module and an ECA module, and the ECA module is sequentially connected to an average pooling module and a decoder module.

[0010] Preferably, the specific workflow of the neural network model in step S2 includes: S2.1 extracts and encodes the input time series data to obtain the encoder module output; S2.2 performs a dimensionality increase operation on the encoder module output, processes it through the ECA module, and then performs dimensionality reduction to obtain the ECA module output; S2.3 performs sequential pooling operation and integration on the ECA module outputs to obtain the average pooling module output; S2.4 performs transposed convolution, batch normalization, and 1D convolution operations on the output of the average pooling module to obtain the output of the decoder module.

[0011] Preferably, the step S2.1 specifically includes: performing 1D convolution processing on the input time series data to obtain the convolution processing output; performing batch normalization processing on the output of the convolution processing to obtain the normalized processing output; performing inactivation processing on the normalized processing output to obtain the inactivation processing output; performing maximum pooling processing and average pooling processing on the inactivation processing output respectively to obtain the maximum pooling processing output and the average pooling processing output; adding the maximum pooling processing output and the average pooling processing output to obtain the output of the encoder module.

[0012] Preferably, the output of the encoder module continues to proceed to step S2.1 until the output of the encoder module is obtained for the fifth time.

[0013] Preferably, the step S2.3 specifically includes: performing 7 time series pooling operations of different scales on the ECA module output to obtain 7 time series pooling outputs, and then integrating the ECA module output and the 7 time series pooling outputs to obtain the output of the average pooling module.

[0014] Preferably, the encoder module includes a 1D convolution layer and a batch normalization layer, and also includes a deactivation layer, a maximum pooling layer and an average pooling layer arranged in sequence.

[0015] Preferably, the ECA module further includes a GAP layer, a 1D convolution layer and a Sigmoid function layer which are arranged in sequence.

[0016] The mathematical logic of the ECA module is as follows: 1. Input a convolutional block 2. After GAP, the time series data is compressed to 1×1×C; 3. After removing the dimension 1×1, we get the clustered features without dimensionality reduction 4. Channel weight is calculated as follows: ω=σ(C1D k (y)) Among them, ω is the channel weight, σ is the Sigmoid function, C1D represents a one-dimensional convolution, and this module has only parameter k. Since the ECA module aims to properly capture local cross-channel interactions, it is necessary to determine the coverage of the interaction (i.e., the kernel size k of the 1D convolution). Since it is reasonable that k is proportional to the channel dimension C, there may be a mapping between k and C: C=φ(k) The simplest mapping form is a linear function, which is in the form of φ(k) = γ·kb. However, the mapping range that can be described by this linear relationship is limited. In deep learning, the channel dimension C is usually set to a power of 2. In order to enrich the mapping possibilities, this paper considers extending this linear function to a nonlinear function. By introducing nonlinear elements and exploring more complex mapping relationships, it is possible to solve problems that cannot be solved by linear functions. Therefore, Formula 2 is modified as follows: C=φ(k)=2 (γ·k-b) Then, given the channel dimension C, the kernel size k can be adaptively determined as: where |t| odd represents the nearest odd number of t. In this paper, we set γ and b to 2 and 1, respectively, in all experiments. Obviously, by mapping ψ, high-dimensional channels have longer distance interactions, while low-dimensional channels experience shorter distance interactions by using nonlinear mapping.

[0017] 5. After increasing the dimension 1×1, we get a weight of 1×1×C; 6. Multiply the weight by the input X to get the weighted time series data.

[0018] Preferably, the average pooling layer in the average pooling module performs average pooling processing and 1D convolution processing, as well as batch normalization processing and inactivation processing in sequence.

[0019] Preferably, the step S4 specifically includes inputting the load forecast of a single electrical appliance into the distributed photovoltaic storage microgrid system, and the control center of the distributed photovoltaic storage microgrid system performs energy scheduling according to the load forecast of the single electrical appliance, and ensures uninterrupted power supply to key loads and realizes low-price grid charging according to the forecast results.

[0020] The beneficial effects of the present invention are as follows: in the field of distributed photovoltaic storage microgrid systems, the ECA module is applied to the decomposition model of non-intrusive load monitoring NILM, and the load of a single household appliance is extracted based on the method of local cross-channel interactive load decomposition, so as to provide prediction characteristics and scheduling basis for load forecasting through the load situation, thereby helping to achieve more accurate prediction and more effective energy scheduling, which not only solves the contradiction between the performance of the channel attention mechanism and the complexity of the model in the non-intrusive load decomposition model, but also significantly improves the accuracy of load decomposition, and uses the load decomposed by NILM as a feature, which provides a multi-dimensional basis for load forecasting of photovoltaic energy storage systems, ensures the continuous power supply of important loads, and reduces energy utilization costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the technical solution of the present invention is further described in detail below through embodiments and in combination with the accompanying drawings. It should be understood that the specific implementation method described here is only an optimal embodiment of the present invention, which is only used to explain the present invention and does not limit the scope of protection of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0023] Distributed photovoltaic and storage microgrid systems, as energy systems that integrate solar photovoltaic power generation, energy storage, and microgrid technology, are becoming increasingly important in home and commercial environments. Taking the home distributed photovoltaic and storage microgrid system as an example, the main function of the home distributed photovoltaic and storage microgrid system is to reduce the cost of energy utilization. The energy storage equipment is charged through photovoltaic power generation and grid charging. The distributed photovoltaic and storage microgrid system needs to be charged from the grid due to the uncertainty of photovoltaic power generation, and the volatility of grid prices in different time periods will lead to different charging costs. Therefore, how to dispatch energy has become the focus of the distributed photovoltaic and storage microgrid system. Through energy dispatching, not only can the charging cost be reduced, but also the electricity cost can be reduced, ensuring the continuous power supply of important loads. What determines energy dispatch is the load power consumption, so household load forecasting has become a crucial step.

[0024] For household distributed photovoltaic and energy storage microgrid systems, it is difficult to achieve more accurate energy scheduling and continuous power supply for important loads through load forecasting in the entire area. Based on this problem, this paper proposes a method based on local cross-channel interactive load decomposition to extract the load of a single household appliance, thereby providing prediction characteristics and scheduling basis for load forecasting through this load situation, helping to achieve more accurate predictions and more effective energy scheduling solutions.

[0025] Non-intrusive load monitoring identifies and distinguishes the usage of different electrical appliances by analyzing the overall power consumption pattern, thereby estimating and monitoring the energy consumption of a single load. Load decomposition is a key step in NILM. Recent studies have shown that emerging technologies such as deep learning and attention mechanisms have great potential in load decomposition. However, most load decomposition models have problems with dimensionality reduction and computational complexity. Dimensionality reduction leads to the destruction of model weights, and computational complexity leads to increased computing requirements.

[0026] The method strategy proposed in this invention applies the effective channel attention (ECA) module to the decomposition model of non-intrusive load monitoring (NILM) in the field of distributed photovoltaic storage microgrid systems, which not only solves the contradiction between the performance of the channel attention mechanism and the model complexity in the non-intrusive load decomposition model, but also significantly improves the accuracy of load decomposition. The load decomposed by NILM is used as a feature to provide a more dimensional basis for the load prediction of the photovoltaic energy storage system, ensure the continuous power supply of important loads, and reduce the cost of energy utilization.

[0027] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe the operations (or steps) as sequential processes, many of the operations (or steps) therein can be implemented in parallel, concurrently, or simultaneously. In addition, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but can also have additional steps not included in the drawings; the process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0028] The technical solution of the present invention is further specifically described below through embodiments and in conjunction with the accompanying drawings.

[0029] Embodiment: A distributed photovoltaic storage microgrid system optimization method of this embodiment, such as Figure 1 As shown, the following steps are included: S1 constructs a neural network model for non-intrusive load decomposition based on the effective channel attention (ECA) module. An effective channel attention (ECA) module based on the principle of local cross-channel interaction is constructed, and a neural network model for non-intrusive load decomposition based on the ECA module is constructed. The neural network model for non-intrusive load decomposition includes an encoder module, an ECA module, an average pooling module, and a decoder module. The encoder module includes a 1D convolution layer, a batch normalization layer, a deactivation layer, a maximum pooling layer, and an average pooling layer. The ECA module includes a global average pooling (GAP) layer, a 1D convolution layer, and a sigmoid function layer. The average pooling module only includes an average pooling layer. The decoder module includes a transposed convolution layer, a batch normalization layer, and a 1D convolution layer. Then, the neural network with the local cross-channel interaction load decomposition model as the core decomposes the total power of the appliances in the household distributed photovoltaic storage microgrid system to obtain the power of a single appliance.

[0030] The encoder module is a sequential step of 1D convolutional layers, batch normalization layers, dropout layers, max pooling layers, and average pooling layers.

[0031] The ECA module is an ordered step of a GAP layer, a 1D convolution layer and a Sigmoid function layer.

[0032] The average pooling layer in the average pooling module is an ordered step of average pooling processing, 1D convolution processing, batch normalization processing and inactivation processing.

[0033] The mathematical logic of the ECA module is as follows: 1. Input a convolutional block 2. After GAP, the time series data is compressed to 1×1×C; 3. After removing the dimension 1×1, we get the clustered features without dimensionality reduction 4. Channel weight is calculated as follows: ω=σ(C1D k (y)) Among them, ω is the channel weight, σ is the Sigmoid function, C1D represents a one-dimensional convolution, and this module has only parameter k. Since the ECA module aims to properly capture local cross-channel interactions, it is necessary to determine the coverage of the interaction (i.e., the kernel size k of the 1D convolution). Since it is reasonable that k is proportional to the channel dimension C, there may be a mapping between k and C: C=φ(k) The simplest mapping form is a linear function, which is in the form of φ(k) = γ·kb. However, the mapping range that can be described by this linear relationship is limited. In deep learning, the channel dimension C is usually set to a power of 2. In order to enrich the mapping possibilities, this paper considers extending this linear function to a nonlinear function. By introducing nonlinear elements and exploring more complex mapping relationships, it is possible to solve problems that cannot be solved by linear functions. Therefore, Formula 2 is modified as follows: C=φ(k)=2 (γ·k-b) Then, given the channel dimension C, the kernel size k can be adaptively determined as: where |t| odd represents the nearest odd number of t. In this paper, we set γ and b to 2 and 1, respectively, in all experiments. Obviously, by mapping ψ, high-dimensional channels have longer distance interactions, while low-dimensional channels experience shorter distance interactions by using nonlinear mapping.

[0034] 5. After increasing the dimension 1×1, we get a weight of 1×1×C; 6. Multiply the weight by the input χ to obtain the weighted time series data.

[0035] The average pooling layer in the average pooling module performs average pooling and 1D convolution, as well as batch normalization and inactivation in sequence.

[0036] S2 uses a neural network model to decompose the household distributed photovoltaic storage microgrid system to obtain the power of a single appliance. The specific workflow of the neural network model includes: S2.1 extracts and encodes the input time series data to obtain the output of the encoder module, specifically including: performing 1D convolution processing on the input time series data to obtain the convolution processing output; performing batch normalization processing on the output of the convolution processing to obtain the normalized processing output; performing inactivation processing on the normalized processing output to obtain the inactivation processing output; performing maximum pooling processing and average pooling processing on the inactivation processing output respectively to obtain the maximum pooling processing output and the average pooling processing output; adding the maximum pooling processing output and the average pooling processing output to obtain the output of the encoder module.

[0037] Continue the process of step S2.1 until the output of the fifth encoder module is obtained.

[0038] The encoder module includes a 1D convolutional layer and a batch normalization layer, and also includes a deactivation layer, a maximum pooling layer, and an average pooling layer arranged in sequence.

[0039] S2.2 performs a dimensionality increase operation on the encoder module output, processes it through the ECA module, and then performs dimensionality reduction to obtain the ECA module output; the ECA module also includes a GAP layer, a 1D convolution layer, and a Sigmoid function layer that are arranged in sequence.

[0040] S2.3 performs temporal pooling operations and integration on the output of the ECA module to obtain the output of the average pooling module, specifically including: performing temporal pooling operations of 7 different scales on the output of the ECA module to obtain 7 temporal pooling outputs, and then integrating the ECA module output and the 7 temporal pooling outputs to obtain the output of the average pooling module. The neural network model includes an encoder module and an ECA module, and the ECA module is sequentially connected to the average pooling module and the decoder module.

[0041] S2.4 performs transposed convolution, batch normalization, and 1D convolution operations on the output of the average pooling module to obtain the output of the decoder module.

[0042] S3 uses the decomposed power status of a single appliance as a feature to obtain the load forecast of a single appliance, and provides the predicted load status to the control center of the distributed photovoltaic storage microgrid system for charging and discharging scheduling.

[0043] S4 performs charging and discharging scheduling based on the predicted load conditions. The load forecast of a single appliance is input into the distributed photovoltaic storage microgrid system. The control center of the distributed photovoltaic storage microgrid system performs energy scheduling based on the load forecast of a single appliance, and ensures uninterrupted power supply to key loads and realizes low-price grid charging based on the forecast results.

[0044] In the field of distributed photovoltaic and energy storage microgrid systems, the ECA module is applied to the decomposition model of non-intrusive load monitoring NILM. The load of a single household appliance is extracted based on the method of local cross-channel interactive load decomposition, so as to provide prediction characteristics and scheduling basis for load forecasting through the load situation, helping to achieve more accurate prediction and more effective energy scheduling. It not only solves the contradiction between the performance of the channel attention mechanism and the complexity of the model in the non-intrusive load decomposition model, but also significantly improves the accuracy of load decomposition. The load decomposed by NILM is used as a feature, which provides a more dimensional basis for load forecasting of photovoltaic energy storage systems, ensures the continuous power supply of important loads, and reduces energy utilization costs.

[0045] Example A distributed photovoltaic storage microgrid system optimization method is disclosed. In the field of distributed photovoltaic storage microgrid systems, an ECA module is applied to a decomposition model of non-intrusive load monitoring (NILM), which not only solves the contradiction between the performance of the channel attention mechanism and the model complexity in the non-intrusive load decomposition model, but also significantly improves the accuracy of load decomposition. The method includes the following steps: Step 1, constructing an effective channel attention (ECA) module based on the principle of local cross-channel interaction, and constructing a neural network model of non-intrusive load decomposition based on the ECA module. The neural network model of non-intrusive load decomposition It includes an encoder module, an ECA module, an average pooling module and a decoder module. The encoder module includes a 1D convolution layer, a batch normalization layer, a deactivation layer, a maximum pooling layer and an average pooling layer. The ECA module includes a global average pooling (GAP) layer, a 1D convolution layer and a Sigmoid function layer. The average pooling module only includes an average pooling layer. The decoder module includes a transposed convolution layer, a batch normalization layer and a 1D convolution layer. Then, a neural network with a local cross-channel interactive load decomposition model as the core decomposes the total power of electrical appliances in the household distributed photovoltaic storage microgrid system to obtain the power situation of a single appliance.

[0046] The workflow of the neural network model for non-intrusive load decomposition includes the following steps: S1, extracts and encodes the input time series data, including: S1.1, input time series data for 1D convolution processing to obtain convolution processing output; S1.2, performing batch normalization processing on the output of the convolution processing to obtain a normalized output; S1.3, performing inactivation processing on the normalized output to obtain an inactivation processed output; S1.4, performing maximum pooling processing and average pooling processing on the output of the deactivation processing respectively, to obtain the output of the maximum pooling processing and the output of the average pooling processing; S1.5, adding the output of the maximum pooling process and the output of the average pooling process to obtain the output of the encoder module; S1.6, continue the steps S1.1 to S1.5 for the output of the encoder module until the output of the fifth encoder module is obtained; S2, the output of the fifth encoder module is subjected to a dimensionality increase operation, processed by the ECA module, and then dimensionality reduction to obtain the ECA module output; S3, perform 7 different scales of time series pooling operations on the output of the ECA module to obtain 7 time series pooling outputs; integrate the output of the ECA module and the 7 time series pooling outputs to obtain the output of the average pooling module; S4, performs transposed convolution, batch normalization and 1D convolution operations on the output of the average pooling module to obtain the output of the decoder module.

[0047] The encoder module is a sequential step of 1D convolutional layer, batch normalization layer, dropout layer, max pooling layer, and average pooling layer.

[0048] Among them, the ECA module is an ordered step of GAP layer, 1D convolution layer and Sigmoid function layer.

[0049] Among them, the average pooling layer in the average pooling module is an ordered step of average pooling processing, 1D convolution processing, batch normalization processing and inactivation processing.

[0050] Among them, the mathematical logic of the ECA module is as follows: S1, input a convolutional block S2, compressed time series data after GAP to 1 × 1 × C; S3, after removing the dimension 1×1, we get the clustered features without dimensionality reduction S4, channel weights are calculated as follows: ω=σ(C1D k (y)) Among them, ω is the channel weight, σ is the Sigmoid function, C1D represents a one-dimensional convolution, and this module has only parameter k. Since the ECA module aims to properly capture local cross-channel interactions, it is necessary to determine the coverage of the interaction (i.e., the kernel size k of the 1D convolution). Since it is reasonable that k is proportional to the channel dimension C, there may be a mapping between k and C: C=φ(k) The simplest mapping form is a linear function, which is in the form of φ(k) = γ·kb. However, the mapping range that can be described by this linear relationship is limited. In deep learning, the channel dimension C is usually set to a power of 2. In order to enrich the mapping possibilities, this paper considers extending this linear function to a nonlinear function. By introducing nonlinear elements and exploring more complex mapping relationships, it is possible to solve problems that cannot be solved by linear functions. Therefore, Formula 2 is modified to: C = φ(k) = 2 (γ·k-b) Then, given the channel dimension C, the kernel size k can be adaptively determined as: where |t| oddrepresents the nearest odd number of t. In this paper, we set γ and b to 2 and 1, respectively, in all experiments. Obviously, by mapping ψ, high-dimensional channels have longer distance interactions, while low-dimensional channels experience shorter distance interactions by using nonlinear mapping.

[0051] S5, after increasing the dimension 1×1, we get a weight of 1×1×C; S6, multiply the weight by the input X to obtain the weighted time series data.

[0052] Step 2, providing the decomposed power situation of a single electrical appliance as a feature to a load prediction neural network model in a distributed photovoltaic storage microgrid system to obtain a load prediction of a single electrical appliance; Among them, the characteristic parameters of the load prediction neural network model provided to the distributed photovoltaic storage microgrid system include the power status of a single appliance decomposed by NILM.

[0053] Step 3, providing the predicted load conditions to the distributed photovoltaic storage microgrid system control center for charging and discharging scheduling; wherein, the distributed photovoltaic storage microgrid system control center performs energy scheduling based on the load forecast of a single electrical appliance, and ensures uninterrupted power supply to key loads and realizes low-price grid charging based on the forecast results.

[0054] Compared with the prior art, the present invention has the following beneficial effects: (1) The decomposed electrical load characteristics provide a more dimensional basis for load prediction of photovoltaic energy storage systems; (2) It provides new ideas for ensuring continuous power supply of important loads and reducing energy utilization costs; (3) Solved the contradiction between the performance of the channel attention mechanism and the model complexity in the non-intrusive load decomposition model; (4)Significantly improved the accuracy of load decomposition.

[0055] (5) Taking the load decomposed by NILM as a feature provides a more dimensional basis for load forecasting of the photovoltaic energy storage system, ensuring the continuous power supply of important loads and reducing the cost of energy utilization.

[0056] The specific embodiments described herein are merely examples of the spirit of the present invention. The above embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that technicians in the technical field to which the present invention belongs can make various modifications or supplements to the described specific embodiments or replace them in a similar manner, but will not deviate from the spirit of the present invention or exceed the scope defined in the attached claims. For ordinary technicians in this field, multiple deformations and improvements can also be made without departing from the concept of the present invention. Therefore, the scope of protection of the patent of the present invention shall be based on the attached claims.

Claims

1. A distributed photovoltaic storage microgrid system optimization method, characterized in that: The following steps are involved: S1 builds a neural network model for non-intrusive load decomposition based on the ECA module; S2 uses a neural network model to decompose the household distributed photovoltaic storage microgrid system to obtain the power of a single appliance; S3 obtains the load forecast of a single appliance by taking the decomposed power situation of the single appliance as the feature; S4 performs charging and discharging scheduling according to the predicted load conditions.

2. A distributed photovoltaic storage microgrid system optimization method according to claim 1, characterized in that: The neural network model in step S1 includes an encoder module and an ECA module, and the ECA module is sequentially connected to an average pooling module and a decoder module.

3. A distributed photovoltaic storage microgrid system optimization method according to claim 2, characterized in that: The specific workflow of the neural network model in step S2 includes: S2.1 extracts and encodes the input time series data to obtain the encoder module output; S2.2 performs a dimensionality increase operation on the encoder module output, processes it through the ECA module, and then performs dimensionality reduction to obtain the ECA module output; S2.3 performs sequential pooling operation and integration on the ECA module outputs to obtain the average pooling module output; S2.4 performs transposed convolution, batch normalization, and 1D convolution operations on the output of the average pooling module to obtain the output of the decoder module.

4. A distributed photovoltaic storage microgrid system optimization method according to claim 3, characterized in that: The step S2.1 specifically includes: performing 1D convolution processing on the input time series data to obtain the convolution processing output; performing batch normalization processing on the output of the convolution processing to obtain the normalized processing output; performing inactivation processing on the normalized processing output to obtain the inactivation processing output; performing maximum pooling processing and average pooling processing on the inactivation processing output respectively to obtain the maximum pooling processing output and the average pooling processing output; adding the maximum pooling processing output and the average pooling processing output to obtain the output of the encoder module.

5. A distributed photovoltaic storage microgrid system optimization method according to claim 4, characterized in that: The output of the encoder module continues the process of step S2.1 until the output of the fifth encoder module is obtained.

6. A distributed photovoltaic storage microgrid system optimization method according to claim 3, 4 or 5, characterized in that: The step S2.3 specifically includes: performing 7 time series pooling operations of different scales on the output of the ECA module to obtain 7 time series pooling outputs, and then integrating the ECA module output and the 7 time series pooling outputs to obtain the output of the average pooling module.

7. A distributed photovoltaic storage microgrid system optimization method according to claim 2, characterized in that: The encoder module includes a 1D convolution layer and a batch normalization layer, and also includes a deactivation layer, a maximum pooling layer and an average pooling layer arranged in sequence.

8. A distributed photovoltaic storage microgrid system optimization method according to claim 2, characterized in that: The ECA module also includes a GAP layer, a 1D convolution layer and a Sigmoid function layer which are arranged in sequence.

9. A distributed photovoltaic storage microgrid system optimization method according to claim 2, characterized in that: The average pooling layer in the average pooling module performs average pooling processing and 1D convolution processing, as well as batch normalization processing and inactivation processing in sequence.

10. A distributed photovoltaic storage microgrid system optimization method according to claim 1 or 2, characterized in that: The step S4 specifically includes inputting the load forecast of a single appliance into the distributed photovoltaic and energy storage microgrid system. The control center of the distributed photovoltaic and energy storage microgrid system performs energy dispatch according to the load forecast of the single appliance, and ensures uninterrupted power supply to key loads and realizes low-price grid charging according to the forecast results.

Images