Non-intrusive renewable energy load prediction system and method based on double convolutional neural networks

Abstract

The invention discloses a non-intrusive renewable energy load prediction system and method based on a double convolutional neural network, and the method comprises the steps: determining an electric equipment state change event through a sliding window event detection algorithm, calculating a power difference before and after the event according to the electric equipment state change event, and taking the power difference before and after the event as a load feature; extracting time sequence features and spatial features in the load features by using a state convolutional neural network, generating predicted power state distribution according to the time sequence features and the spatial features, inputting the predicted power state distribution into a power convolutional neural network, and calculating power consumption under the predicted power state distribution; and converting the predicted power state distribution and power consumption into the working state of each electric device at each sampling time point, and generating a monitoring report according to the working state. According to the invention, the working state of each electric device can be monitored and analyzed without directly accessing each electric device, and meanwhile, the load monitoring precision can be improved and the equipment investment cost can be reduced.

Description

Technical Field

[0001] The present invention relates to the field of renewable energy power generation control technology, and specifically to a non-invasive renewable energy load forecasting system and method based on a dual convolutional neural network, and in particular to a non-invasive renewable energy load forecasting system and method based on a dual convolutional neural network combined with renewable energy power generation. Background Art

[0002] For systems that are connected to photovoltaic, energy storage, and power loads within the microgrid, by monitoring the loads and understanding the load characteristics of the power equipment, it is possible to assist in the dispatching and control of new energy resources such as photovoltaics and energy storage. Traditional load monitoring methods often require direct access to each power equipment, which not only increases the complexity of installation and maintenance, but may also infringe on user privacy. Non-intrusive load monitoring (NILM) technology only collects the user's total power consumption data and uses algorithms to decompose the energy consumption information of each power equipment, thereby realizing the monitoring and analysis of the user's power consumption behavior. Non-intrusive load monitoring can be divided into several types of methods, including those based on steady-state characteristics (including power characteristics and current harmonic characteristics), transient characteristics (including transient current characteristics and transient power characteristics), event detection (including event triggering and pattern recognition), and machine learning (including supervised learning and unsupervised learning). Among them, the steady-state feature method is relatively simple to calculate but difficult to distinguish; the transient feature method has high requirements for equipment and algorithms but has high recognition; the event detection method can respond in a timely manner but is easily interfered with; the machine learning method has high accuracy and generalization ability but supervised learning requires a large amount of labeled data and has high computational costs, and the accuracy of unsupervised learning is relatively low. Summary of the invention

[0003] The purpose of the present invention is to solve the problem that each method of non-intrusive load monitoring in the prior art has certain defects, and thus proposes a non-intrusive renewable energy load forecasting system and method based on a dual convolutional neural network. The present invention collects the photovoltaic power generation of the microgrid and the total power consumption data of the user's mains electricity, without directly accessing each power-consuming device, so as to realize the monitoring and analysis of the working status of each power-consuming device, and at the same time can improve the load monitoring accuracy and reduce the equipment investment cost.

[0004] To achieve this purpose, the present invention designs a non-intrusive renewable energy load forecasting system based on a dual convolutional neural network, which includes a load feature calculation module, a power state distribution and power consumption information acquisition module, and a result output module;

[0005] The load characteristic calculation module determines the state change event of the power consumption device through the sliding window event detection algorithm, and calculates the power difference before and after the event according to the state change event of the power consumption device, and uses the power difference before and after the event as the load characteristic;

[0006] The power state distribution and power consumption information acquisition module uses a state convolutional neural network to extract the time series characteristics and spatial characteristics in the load characteristics, generates a predicted power state distribution according to the time series characteristics and spatial characteristics, and inputs the predicted power state distribution into the power convolutional neural network to calculate the power consumption of the electrical equipment under the predicted power state distribution;

[0007] The result output module is used to convert the predicted power state distribution and the power consumption of the electric devices under the predicted power state distribution into the working state of each electric device at each sampling time point, and generate a monitoring report.

[0008] Beneficial effects of the present invention:

[0009] The present invention firstly uses an innovative sliding window event detection algorithm to accurately capture the state change events of electrical equipment and accurately determine the starting and ending points of the events. This method based on the change of power mean and variance effectively improves the recognition accuracy of the switching state of electrical equipment. At the same time, the dynamic time warping (DTW) algorithm is introduced to analyze the event set, which realizes the nonlinear time alignment of the operation modes of complex electrical equipment, further improves the matching degree and recognition accuracy between events of different electrical equipment, and finally adopts a convolutional neural network architecture (including SCN and PCN) to more accurately capture the power state distribution and power consumption information of electrical equipment, thereby improving the overall monitoring accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a structural block diagram of the non-intrusive renewable energy load forecasting system based on dual convolutional neural networks of the present invention;

[0011] Figure 2 Schematic diagram of event monitoring based on sliding window in the present invention;

[0012] Figure 3 The load diagram of five types of typical electrical equipment in Example 3 of the present invention;

[0013] Figure 4 This is a working status diagram of five types of typical electrical equipment at various sampling time points in Example 3 of the present invention. DETAILED DESCRIPTION

[0014] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0015] Example 1

[0016] A non-intrusive renewable energy load forecasting system based on dual convolutional neural networks, such as Figure 1 As shown, it includes a load characteristic calculation module, a power state distribution and power consumption information acquisition module, and a result output module;

[0017] The load characteristic calculation module determines the state change event of the power consumption device through the sliding window event detection algorithm, and calculates the power difference before and after the event according to the state change event of the power consumption device, and uses the power difference before and after the event as the load characteristic;

[0018] The power state distribution and power consumption information acquisition module uses a state convolutional neural network to extract the time series characteristics and spatial characteristics in the load characteristics, generates a predicted power state distribution according to the time series characteristics and spatial characteristics, and inputs the predicted power state distribution into the power convolutional neural network to calculate the power consumption of the electrical equipment under the predicted power state distribution;

[0019] The result output module is used to convert the predicted power state distribution and the power consumption of the electric devices under the predicted power state distribution into the working state of each electric device at each sampling time point, and generate a monitoring report based on the working state.

[0020] In the above technical solution, the system also includes a data acquisition and preprocessing module, which is used to collect the user's total power consumption data, including the voltage and current of each power-consuming device. In the specific operation process, the voltage and current need to be synchronously recorded with a timestamp for subsequent event detection and feature extraction, and the voltage and current need to be cleaned to remove noise and abnormal values, and the cleaned voltage and current are converted into active power and reactive power. Specifically, an oscilloscope DL750 is used as a data acquisition device to sample the voltage and current at the bus at a frequency of 1kHz. This high-frequency sampling can ensure that subtle fluctuations when the state of the power-consuming device changes are captured.

[0021] In the above technical solution, the specific method of determining the state change event of the electric device through the sliding window event detection algorithm is as follows: Figure 2 As shown in FIG. 1 , the relationship curve between the sampling point and the total power consumption of the user is used as the detection range of the state change event of the power-consuming device. A sliding window with a total length of 2N is set on the relationship curve. The sliding window is divided into two sub-windows of equal length, which can be expressed as W = [P1, ..., P N ,P N+1 ,…,P 2N ], when the state of the power-consuming equipment changes, the mean and variance of the power will gradually change, and this is used as a basis to find the starting and ending points of the power-consuming equipment state change event. When the sliding window transitions from the steady state area to the transient state area, the calculation formula for the power mean in the sliding window is as follows:

[0022]

[0023] Where E1 is the average power in the sliding window from the steady state region to the transient state region, P nis the power of the nth sampling point, and N is the number of sampling points;

[0024] When the window of the sliding window transitions from the steady state region to the transient region, the calculation formula for the power variance within the sliding window is as follows:

[0025]

[0026] In the formula, D1 is the power variance of the 1st to the Nth sampling points within the sliding window from the steady state region to the transient region, E1 is the average power within the sliding window from the steady state region to the transient region, P n is the power of the nth sampling point, and N is the number of sampling points;

[0027]

[0028] In the formula, D2 is the power variance of the (N + 1)th to the 2Nth sampling points within the sliding window from the steady state region to the transient region, E1 is the average power within the sliding window from the steady state region to the transient region, P n is the power of the nth sampling point, and N is the number of sampling points;

[0029] When D2 - D1 > T1, it is determined that an event of the state change of the electrical equipment occurs. Starting from the 2Nth sampling point of the sliding window, traverse the sampling points to the left. When |P (x) - E1| < T2E1, where P (x) represents the actual power of the electrical equipment at the xth sampling point, and both T1 and T2 are set values, then the sampling point X is the starting point of the event of the state change of the electrical equipment;

[0030] When the window of the sliding window transitions from the transient region to the steady state region, the calculation formula for the average power within the sliding window is as follows:

[0031]

[0032] In the formula, E2 is the average power within the sliding window from the transient region to the steady state region, P n is the power of the nth sampling point, and N is the number of sampling points;

[0033] When the window of the sliding window transitions from the transient region to the steady state region, the calculation formula for the power variance within the sliding window is as follows:

[0034]

[0035] In the formula, D1′ is the power variance of the 1st to the Nth sampling points within the sliding window from the transient region to the steady state region, E2 is the average power within the sliding window from the transient region to the steady state region, P n is the power of the nth sampling point, and N is the number of sampling points;

[0036]

[0037] Wherein, D2' is the power variance of the (N + 1)-th to 2N-th sampling points in the sliding window during the transition from the transient region to the steady state region, E2 is the power mean value in the sliding window during the transition from the transient region to the steady state region, P n is the power at the n-th sampling point, and N is the number of sampling points;

[0038] When D1' - D2' < T3, it is determined that the event of the change in the state of the electrical equipment has ended. Starting from the first sampling point of the sliding window, traverse the sampling points to the right. When |P (n) - E2| < T4E2, where P (n) represents the power at the sampling point n, and both T3 and T4 are set values, then the sampling point n is the end point of the event of the change in the state of the electrical equipment. Through the innovative sliding window event detection algorithm, the event of the change in the state of the electrical equipment can be accurately captured, and the start point and end point of the event can be accurately determined. This method based on the change of power mean value and variance effectively improves the recognition accuracy of the switch state of the electrical equipment

[0039] In this article, the transient region refers to the time period during which the total power consumption is in an unstable state during the change of the state of the electrical equipment. When the state of the electrical equipment changes, such as turning on or off, the power will experience a transition process from the old state to the new state. During this transition process, the mean value and variance of the power will gradually change, and the state of the system at this time is the transient region. For example, when an air conditioner starts, the power will rise rapidly from a relatively low level, and this process of power rise is in the transient region. The characteristics of the transient region include: large power fluctuations: due to the change of the equipment state, the total power consumption is unstable and the fluctuations are obvious; relatively short duration: the transient process is usually a short transition stage, and once the equipment reaches the new stable state, it will leave the transient region.

[0040] In this article, the steady state region refers to the time period during which the electrical equipment is in a stable operating state and the total power consumption is relatively stable. In the steady state region, the electrical equipment has completed the state change, and the mean value and variance of the power basically remain stable. For example, when the air conditioner is operating stably, the power is maintained at a relatively stable level, and the system is in the steady state region at this time. The characteristics of the steady state region include: relatively stable power: the equipment is operating stably, and the total power consumption fluctuates less; long duration: as long as the equipment maintains a stable operating state, the system will always be in the steady state region until a new event of the change in the equipment state occurs.

[0041] In this article, through a large number of experiments, the reasonable values of the parameters are obtained: N = 20, T1 = D1, T2 = 0.2, T3 = D2', T4 = 0.2. Therefore, in the above technical solution, the calculation formula for the power difference before and after the event calculated according to the event of the change in the state of the electrical equipment is as follows:

[0042]

[0043] Where ΔP is the power difference between the starting point and the end point of the state change event of the electrical equipment, P begin is the power at the start of the power equipment state change event, P end The power difference is the power at the end point of the power consumption device state change event. Specifically, it can be further determined whether the suspicious event represents a state change of the power consumption device according to the power difference.

[0044] In the above technical solution, the system also includes a data matching module, which first records the detected events and their related features into a transient event record library, and uses an association rule mining algorithm to analyze the transient event record library to find out the event set belonging to the same electrical device. In this article, an event refers to the turning on or off of a device, and the transient event record library refers to the power change when the device is turned on or off. The specific operations of using an association rule mining algorithm to analyze the transient event record library and find out the event set belonging to the same electrical device are: 1. Construct a distance matrix: Use the dynamic time warping (DTW) algorithm to calculate the matching degree of the two sets of data. Suppose the two sets of sequence data are X = {x1, x2,…, x n} and Y = {y1, y2, ..., y m}, construct an n×m distance matrix D to align the two sets of sequence data nonlinearly, and the element d at (i, j) in the matrix ij Represents sequence data x i and j The distance between them, then the matrix D can be expressed as the distance between each data point in the sequence X and each data point in the sequence Y, where the matrix D is shown as follows:

[0045]

[0046] 2. Determine the path search range: From the matrix D, we can see that when n=m, the distance matrix D is a square matrix. 11 to d nm The sum of the diagonal distances is the linear alignment match of the two sets of sequence data. When n≠m, the path W={w1,…,w k ,…,w K}, the nonlinear matching degree of two sets of sequence data is calculated from d 11 to d nm The cumulative sum of the curve distance values, where max(n,m)≤K≤(n+m-1). This path must satisfy both monotonicity and continuity. If w k-1 =d ab , then the next point w on the path k =d cdIt must satisfy 0≤(ca)≤1 and 0≤(db)≤1. This ensures that each data point of the sequence data X and Y participates in the calculation in turn, while achieving nonlinear alignment of the time axis, which can be intuitively understood as the shortening and lengthening of the time axis. Therefore, when the kth value w of the path k =d ij When k+1 It can only be D (i+1)j d i(j+1) or (i+1)(j+1) , that is, there are three choices for each step on the path to the next step, and there are 3 paths that meet the above constraints k-1 The matching degree of two sets of sequence data X and Y is defined as:

[0047]

[0048] 3. Dynamic programming calculation: Starting from the starting point, the minimum cumulative distance to each position of the end point is calculated step by step; for each position (i, j), its minimum cumulative distance can be determined by selecting the smallest one of the following three possible paths: from (i-1, j), from (i, j-1) or from (i-1, j-1). The specific calculation formula can be expressed as:

[0049] DW(i,j)=d(X i , Y j )+min{DW(i-1,j),DW(i,j-1),DW(i-1,j-1)} (11)

[0050] 4. Determine the best path and the minimum regularization cost: The dynamic programming algorithm can be used to find the path with the minimum regularization cost. The cumulative distance on this path is Min{DW(X,Y)}, which is the minimum regularization cost. The formula for the best matching function is as follows:

[0051]

[0052] In the formula, DTW (X,Y) represents the dynamic time warping distance between two sets of sequence data X and Y, min{DW (X，Y)} represents the intermediate variable used to calculate the cumulative distance in the dynamic time warping algorithm, X represents the power data sequence of a certain power device at different times, and Y represents the power data sequence of another related power device at different times. The dynamic time warping (DTW) algorithm is introduced to analyze the event set, which realizes the nonlinear time alignment of the operation mode of complex power devices, and further improves the matching degree and recognition accuracy between events of different power devices.

[0053] In the above technical solution, the state convolutional neural network is used to extract the time series features and spatial features in the load features, and the specific method of generating the predicted power state distribution according to the time series features and spatial features is as follows: first, the electrical equipment is divided into k groups according to the power characteristic category, and a classifier is added through the Attention mechanism. The classifier is used to perform sampling weights on the small class sample data in the k groups of electrical equipment, and weighted penalties are imposed on the misclassification of the small class electrical equipment in the k groups of electrical equipment; in this article, the small class sample data refers to the data samples of a relatively small number of electrical equipment in a certain group among the k groups of electrical equipment divided above, and the small class electrical equipment refers to a specific type of electrical equipment that accounts for a small proportion in the entire electrical equipment system. Misclassification refers to the incorrect classification of small class electrical equipment into other categories during the classification process; the specific operation of sampling weights is: through the Attention mechanism, according to the scarcity of small class sample data in the training data and its importance to the accuracy of the overall model, it is given a relatively large weight in the classifier calculation process, and the specific operation of weighted penalties is: when a small class electrical equipment is misclassified, a corresponding penalty term is added to the loss function. The prior art does not specifically consider the case of small-category sample data for different categories of electrical equipment. However, in the state convolutional neural network (SCN) of the present invention, the sampling weight of the classifier for small-category sample data is increased through the Attention mechanism, and weighted penalties are imposed on misclassification of small-category electrical equipment. This improvement allows the model to pay more attention to small-category samples during the training process, thereby improving the recognition accuracy of small-category electrical equipment and avoiding the situation where small-category samples are "overwhelmed" by large-category samples.

[0054] Next, the convolutional layer, pooling layer and fully connected layer in the state convolutional neural network are used to extract the temporal and spatial features of the load features. In the specific operation process, the load features of different types of equipment are divided according to the type of electrical equipment corresponding to the load feature data, and the load features of different types of equipment are processed by 1×1 and 3×3 convolution to generate different feature maps to form r feature maps. Specifically, the 1×1 convolution linearly weights the feature map in the time dimension without changing the size of the time dimension. It can capture local temporal features and reflect the power change trend of electrical equipment in a short time, such as the change of power at adjacent sampling points at the moment of equipment start and stop. The 3×3 convolution can consider more adjacent time point information and can capture the power change mode of electrical equipment in a longer period of time, such as the power fluctuation characteristics of periodic working equipment. After obtaining r feature maps, the Attention operation is performed, and the feature fusion is performed by adding the corresponding elements. Finally, the fused features of the k groups are concat with the load features to obtain the output of the final predicted power state distribution.

[0055] In this article, the state convolutional neural network has a structure that can effectively extract the temporal features and spatial features in the load characteristics, including convolutional layers, pooling layers, and fully connected layers, and improves the processing capabilities of the characteristics of different types of electrical equipment through the Attention mechanism and specific operation procedures. It is a convolutional neural network specially designed to process the load characteristic data in the present invention.

[0056] In this article, the specific method of the Attention operation is to perform weighted processing based on the importance of each element in the feature map after obtaining multiple feature maps. For example, different feature map elements are given different weights according to their relevance to the prediction of the final power state distribution, so that the model pays more attention to the features that have an important impact on the prediction results. The corresponding elements refer to the elements in the same position in different feature maps when performing feature fusion operations. For example, if there are two feature maps A and B, the i-th row and j-th column element in A and the i-th row and j-th column element in B are corresponding elements. The specific method of adding is to sum the corresponding elements according to the conventional numerical addition method. For example, the corresponding elements in the above-mentioned A and B feature maps are added separately to obtain the element values ​​in the new feature map.

[0057] In this article, concat refers to concatenating the features of different groups after fusion operation with the original load features in the feature dimension. For example, if there are k groups of features after the above operation, they are sequentially connected with the original load features in a specific dimension to form a new comprehensive feature for subsequent power state distribution prediction.

[0058] A general convolutional neural network may not have a specific fusion method for feature maps processed by convolution kernels of different sizes. In the present invention, for each of the r divided feature maps, 1×1 and 3×3 convolution operations are performed on each feature map, respectively. After obtaining the r feature maps, an Attention operation is performed, and feature fusion is performed by adding corresponding elements. This fusion method can fully extract feature information of different scales and improve the expressiveness of features. At the same time, the traditional method may not make a specific combination of features of different categories of electrical equipment. In the present invention, the fused features of k groups are concat with the load features to obtain the output of the final predicted power state distribution. This method makes the output result more comprehensively consider the features of different categories of electrical equipment and improves the accuracy of power state distribution prediction.

[0059] In the above technical solution, the predicted power state distribution is input into the power convolutional neural network, and the specific method for calculating the power consumption of the power-consuming equipment under the predicted power state distribution is: the input predicted power state distribution is processed by the power convolutional neural network (PCN), and the PCN has its own network structure, including convolutional layer, pooling layer and fully connected layer, etc., and the power consumption of each power-consuming equipment under the predicted power state distribution is finally calculated by extracting and converting the input data. In this article, the power convolutional neural network refers to a convolutional neural network associated with the state convolutional neural network (SCN) and having a specific function. The power convolutional neural network (PCN) shares some convolutional layer parameters with the SCN, and is mainly used to calculate the power consumption of the power-consuming equipment according to the predicted power state distribution generated by the SCN. In the prior art, different neural networks are usually trained independently and do not share parameters, while the power convolutional neural network (PCN) of the present invention shares some convolutional layer parameters with the state convolutional neural network (SCN). This improvement reduces the amount of calculation and improves the training efficiency and running speed of the model.

[0060] In the above technical solution, the power state distribution and power consumption information acquisition module also includes using historical load data as training data, and optimizing the state convolutional neural network and the power convolutional neural network by minimizing the mean and variance of the predicted power state distribution and the actual power state distribution. During the optimization process, the learning parameters are iteratively updated using the best matching function until the preset conditions are met to obtain the optimal state convolutional neural network and power convolutional neural network models. In this article, minimizing the predicted power state distribution means that the difference between the mean of the predicted power state distribution and the mean of the actual power state distribution is as small as possible, and at the same time, the difference between the variance of the predicted power state distribution and the variance of the actual power state distribution is also as small as possible.

[0061] The historical load data is used as training data, and the specific operation process of optimizing the state convolutional neural network and the power convolutional neural network is carried out by minimizing the mean and variance of the predicted power state distribution and the actual power state distribution. (i) Data preprocessing: 1. First, the collected historical load data is cleaned and normalized. The cleaning operation includes removing noise points and outliers in the data, and the normalization process is to map data of different magnitudes to the same interval, such as normalizing the power data to the [0,1] interval; 2. The preprocessed historical load data is divided into a training set and a validation set according to a certain ratio (80% for training and 20% for validation); (ii) Model initialization: Randomly initialize the weights and bias parameters of the state convolutional neural network (SCN) and the power convolutional neural network (PCN). The initialization parameters determine the initial state of the network, and a random initialization method is used, such as uniform distribution or normal distribution within a certain range. (iii) Forward propagation calculation: 1. Input the historical load data in the training set into the SCN. The SCN processes the input data through its convolutional layer, pooling layer, and fully connected layer structures, extracts the temporal and spatial features in the load characteristics, and generates a predicted power state distribution based on these features; 2. Input the predicted power state distribution output by the SCN into the PCN, and the PCN calculates the power consumption of each electrical device under the predicted power state distribution; (iv) Loss calculation: 1. According to the predicted power state distribution output by the SCN and the corresponding actual power state distribution, calculate their mean and variance respectively; 2. Define the loss function as the difference between the predicted power state distribution and the actual The sum of the differences in the mean and variance of the power state distribution; (V) Back propagation and parameter update: 1. Use the back propagation algorithm to calculate the gradient of the loss function for each parameter (weight and bias) in SCN and PCN. The gradient represents the rate of change of the loss function with the change of the parameter; 2. According to the calculated gradient, the parameters of SCN and PCN are updated using the stochastic gradient descent method; (VI) Model evaluation and iteration: 1. After each training cycle, input the validation set data into the currently trained network (SCN and PCN) and calculate the loss value on the validation set; 2. Determine whether the model has converged based on the validation set loss value. If the validation set loss value no longer decreases significantly in multiple consecutive training cycles, or reaches the preset maximum number of training cycles, stop training; otherwise, continue training for the next training cycle.

[0062] In this article, the specific operation process of iteratively updating the learning parameters using the optimal matching function is as follows: in each iteration, the optimal matching function value corresponding to the data in the training set is first calculated based on the current network parameters. This value reflects the degree of matching between the predicted power state distribution and the actual power state distribution; then, the parameter update direction is determined based on the derivative of the network parameters with respect to the optimal matching function value. If the optimal matching function value is large, it means that the difference between the prediction and the actual is large, and it is necessary to update in the direction of reducing the value; then, according to a certain learning rate, the learning parameters in the network (such as weights and biases, etc.) are updated according to the determined direction, and this cycle is repeated until the preset conditions are met. Specifically, for example, the preset conditions are: 1. Convergence condition: the value of the loss function on the validation set. If the change in the value of the loss function is less than a certain threshold value in several consecutive training cycles, the model is considered to have converged and the training can be stopped; 2. Maximum number of iterations: a maximum number of training cycles is preset, such as 1000 times. When the number of network training cycles reaches this maximum number of iterations, the training is stopped regardless of whether the model converges; 3. Performance indicators meet the requirements: the accuracy of the predicted working state of the electrical equipment on the validation set reaches more than 95%, and the training is stopped. The purpose of optimizing the state convolutional neural network and the power convolutional neural network is to reduce the model training time and thus improve efficiency.

[0063] In the above technical solution, the result output module also includes experimental verification, that is, comparing the monitoring results with the actual situation to verify the accuracy and reliability of the system. The output result of the deep learning model module is the prediction of the working status of each electrical equipment at each sampling time point. The result output module generates a monitoring report based on the output results of the deep learning model to provide users with intuitive load monitoring information. The accuracy and reliability of the deep learning model module directly affect the quality of the result output module, and the result output module presents the results of the deep learning model to the user, achieving the ultimate goal of the entire non-invasive load monitoring system.

[0064] Example 2

[0065] A non-intrusive renewable energy load forecasting method based on a dual convolutional neural network comprises the following steps:

[0066] Determine the state change event of the power consumption device through the sliding window event detection algorithm, and calculate the power difference before and after the event according to the state change event of the power consumption device, and use the power difference before and after the event as the load feature;

[0067] Extracting the time series features and the spatial features in the load features by using a state convolutional neural network, generating a predicted power state distribution according to the time series features and the spatial features, and inputting the predicted power state distribution into the power convolutional neural network to calculate the power consumption under the predicted power state distribution;

[0068] The predicted power state distribution and power consumption are converted into the working state of each electrical device at each sampling time point, and a monitoring report is generated based on the working state.

[0069] Example 3

[0070] In this embodiment, five typical types of electrical equipment in the microgrid are selected: photovoltaic, energy storage, air conditioning, cooling load, and heating load. These electrical equipment are connected to a bus, and the voltage and current are sampled at the bus with a DL750 at a frequency of 1kHz. The characteristics of the power load, heating load, cooling load, and energy storage load are collected by the oscilloscope DL750, as shown in FIG. Figure 3 As shown, from Figure 3 It can be seen that during the start-up process of the energy storage load electrical equipment, the active power increases in a step-like manner, and basically no peak appears, but its reactive power has a high peak. During the shutdown process, the active power decreases in a step-like manner, and the reactive power first rises to the peak and then returns to zero. Other types of electrical equipment also have their own characteristics. The transient waveform is determined by the components and composition of the electrical equipment. Different types of electrical equipment often differ greatly, so the power transient waveform has a good resolution and can be used as a load feature for identification. Common loads in microgrids can be divided into linear loads and nonlinear loads. The power-consuming elements of linear loads (air conditioners) are pure resistors, so their transient process current fluctuations are smooth and the harmonic content is low. Nonlinear loads (energy storage) include inductive loads and capacitive loads. The switching of such loads may be accompanied by impact currents, and it takes a long time to return to a steady state.

[0071] Data acquisition and preprocessing module: Four typical types of electrical equipment in the microgrid are selected - energy storage, air conditioning, cold load equipment, and heat load equipment. These electrical equipment are connected through a bus to ensure that their total power consumption data can be collected at the same time. The oscilloscope DL750 is used as a data acquisition device to sample the voltage and current at the bus at a frequency of 1kHz. The collected raw data is cleaned to remove noise and abnormal values. At the same time, the voltage and current data are converted into active power and reactive power.

[0072] Load feature calculation module: According to the characteristics of the state change of the power-consuming equipment, a sliding window with equal length is set (the total length is set according to the optimal length determined by the experiment), and the mean and variance of the power in the sliding window are calculated in real time. When the change of the mean and variance exceeds the preset threshold, it is determined that the power-consuming equipment state change event has occurred. The sliding window is used to traverse the sample points forward and backward to find the starting point and end point of the event respectively. Then, the power difference before and after the event is calculated according to formula (7) as the load feature;

[0073] Data matching module: records the detected events and their related features into the transient event record library, uses the association rule mining algorithm to analyze the transient event record library, and finds the event set belonging to the same electrical equipment. Specifically, the DTW algorithm is used to process the event set of the same electrical equipment and the matching degree between the events is calculated;

[0074] Power state distribution and power consumption information acquisition module: Based on the architecture of SCN and PCN, a convolutional neural network model is constructed. Historical load data is used as training data. The model is trained by minimizing the mean and variance of the predicted power state distribution and the actual power state distribution. During the training process, the learning parameters are iteratively updated using the best matching function. The model performance is optimized by adjusting the network structure, learning rate, optimizer and other parameters of SCN and PCN until the preset conditions are met and the optimal model is obtained. SCN is used to capture the power state distribution, and PCN is used to predict power consumption information.

[0075] Result output module: converts the output results of the deep learning model into the working status of each electrical equipment at each sampling time point. The results are as follows: Figure 4 As shown, finally, the monitoring results are compared with the actual situation to verify the accuracy and reliability of the system.

[0076] Example 4

[0077] A computer-readable storage medium stores a computer program, wherein the computer program implements the steps of the above method when executed by a processor.

[0078] The present invention determines the state change event of the electrical equipment through a sliding window event detection algorithm, and calculates the power difference before and after the event as the load feature. These load features provide input data for the deep learning model module. At the same time, the transient event record library is analyzed through the association rule mining algorithm and the dynamic time warping (DTW) algorithm to find the event set belonging to the same electrical equipment, and perform nonlinear time alignment to improve the recognition accuracy. These analysis results can also provide more accurate event information and feature data for the deep learning model module, helping the model to better capture the power state distribution and power consumption sequence matching of the electrical equipment; the output result of the deep learning model module is the working state prediction of each electrical equipment at each sampling time point. According to the output result of the deep learning model, a monitoring report is generated to provide users with intuitive load monitoring information; the accuracy and reliability of the deep learning model module directly affect the quality of the result output module, and the result output module presents the results of the deep learning model to the user, achieving the ultimate goal of the entire non-intrusive load monitoring system.

[0079] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A non-intrusive renewable energy load forecasting system based on dual convolutional neural networks, characterized by: It includes a load characteristic calculation module, a power state distribution and power consumption information acquisition module, and a result output module; The load characteristic calculation module determines the state change event of the power consumption device through the sliding window event detection algorithm, and calculates the power difference before and after the event according to the state change event of the power consumption device, and uses the power difference before and after the event as the load characteristic; The power state distribution and power consumption information acquisition module uses a state convolutional neural network to extract the time series characteristics and spatial characteristics in the load characteristics, generates a predicted power state distribution according to the time series characteristics and spatial characteristics, and inputs the predicted power state distribution into the power convolutional neural network to calculate the power consumption of the electrical equipment under the predicted power state distribution; The result output module is used to convert the predicted power state distribution and the power consumption of the electric devices under the predicted power state distribution into the working state of each electric device at each sampling time point, and generate a monitoring report based on the working state.

2. The renewable energy load forecasting system according to claim 1, characterized in that: The specific method for determining the state change event of the electric device by the sliding window event detection algorithm is as follows: taking the relationship curve between the sampling point and the total power consumption of the user as the detection range of the state change event of the electric device, and setting a sliding window with a total length of 2N on the relationship curve; When the sliding window transitions from the steady state region to the transient state region, the calculation formula for the power mean in the sliding window is as follows: Where E1 is the average power in the sliding window from the steady state region to the transient state region, P n is the power of the nth sampling point, N is the number of sampling points; When the sliding window transitions from the steady state region to the transient state region, the calculation formula of the power variance in the sliding window is as follows: Where D1 is the power variance of the 1st to Nth sampling points in the sliding window from the steady state region to the transient region, E1 is the power mean in the sliding window from the steady state region to the transient region, P n is the power of the nth sampling point, N is the number of sampling points; Where D2 is the power variance of the N+1th to 2Nth sampling points in the sliding window from the steady state region to the transient region, E1 is the power mean in the sliding window from the steady state region to the transient region, and P n is the power of the nth sampling point, N is the number of sampling points; When D2 - D1 > T1, it is determined that a useful electrical equipment state change event has occurred. Starting from the 2Nth sampling point of the sliding window and traversing the sampling points to the left, when |P (x) - E1| < T2E1, where P (x) represents the power at sampling point X, and both T1 and T2 are set values, then sampling point X is the starting point of the electrical equipment state change event; When the sliding window transitions from the transient region to the steady-state region, the calculation formula for the power mean in the sliding window is as follows: Where E2 is the average power in the sliding window from the transient region to the steady state region, P n is the power of the nth sampling point, N is the number of sampling points; When the sliding window transitions from the transient region to the steady-state region, the calculation formula for the power variance in the sliding window is as follows: Where D1′ is the power variance of the 1st to Nth sampling points in the sliding window from the transient region to the steady state region, E2 is the power mean in the sliding window from the transient region to the steady state region, and P n is the power of the nth sampling point, N is the number of sampling points; Where D2′ is the power variance of the N+1th to 2Nth sampling points in the sliding window from the transient region to the steady state region, E2 is the power mean in the sliding window from the transient region to the steady state region, and P n is the power of the nth sampling point, N is the number of sampling points; When D1′ - D2′ < T3, it is determined that the electrical equipment state change event has ended, and the sampling points are traversed starting from the first sampling point of the sliding window to the right. When |P (n) - E2| < T4E2, where P (n) represents the power at sampling point n, and both T3 and T4 are set values, then sampling point n is the end point of the electrical equipment state change event.

3. The renewable energy load forecasting system according to claim 2, characterized in that: The formula for calculating the power difference before and after the event based on the state change event of the power-consuming equipment is as follows: Where ΔP is the power difference between the starting point and the end point of the state change event of the electrical equipment, P begin is the power at the start of the power equipment state change event, P end It is the power at the end point of the power-consuming device state change event.

4. The renewable energy load forecasting system according to claim 3, characterized in that: The specific method of extracting the time series features and spatial features in the load features by using the state convolutional neural network and generating the predicted power state distribution according to the time series features and spatial features is as follows: the electrical equipment is divided into several groups according to the power characteristic categories, a classifier is added through the Attention mechanism, the classifier is used to perform sampling weights on the small class sample data in the several groups of electrical equipment, and a weighted penalty is performed on the misclassification of the small class electrical equipment in the several groups of electrical equipment; The convolution layer, pooling layer and fully connected layer in the state convolutional neural network are used to extract the temporal features and spatial features in the load features, and the temporal features and spatial features are subjected to 1×1 and 3×3 convolution operations respectively to obtain the convolved feature map, and the convolved feature map is subjected to Attention operation. The feature maps after the Attention operation in several groups of electrical equipment are subjected to feature fusion by adding corresponding elements, and the fused features are connected with the load features to obtain the predicted power state distribution.

5. The renewable energy load forecasting system according to claim 4, characterized in that: The predicted power state distribution is input into the power convolutional neural network, and the specific method for calculating the power consumption of the electrical equipment under the predicted power state distribution is as follows: first, the local features in the predicted power state distribution are extracted by the convolutional layer in the power convolutional neural network, and then the local features are reduced in dimension by the pooling layer in the power convolutional neural network to retain the main features, and then the fully connected layer in the power convolutional neural network integrates the mapping features, and the power consumption of the electrical equipment under the corresponding predicted power state distribution is obtained in the output layer.

6. The renewable energy load forecasting system according to claim 5, characterized in that: The system also includes a data matching module, which uses a dynamic time warping algorithm to find the path with the minimum warping cost for two sets of sequence data in the transient event record library, and obtains the formula of the best matching function as shown below: DTW (X,Y) =min{DW (X,Y) } (8) In the formula, DTW (X,Y) represents the dynamic time warping distance between two sets of sequence data X and Y, min{DW (X，Y) } represents the intermediate variable used to calculate the cumulative distance in the dynamic time warping algorithm, X represents the power data sequence of a certain electrical equipment at different times, and Y represents the power data sequence of another related electrical equipment at different times.

7. The renewable energy load forecasting system according to claim 6, characterized in that: The power state distribution and power consumption information acquisition module also includes using historical load data as training data, and optimizing the state convolutional neural network and the power convolutional neural network by minimizing the mean and variance of the predicted power state distribution and the actual power state distribution. During the optimization process, the learning parameters are iteratively updated using the best matching function until the preset conditions are met, thereby obtaining the optimal state convolutional neural network and power convolutional neural network models.

8. The renewable energy load forecasting system according to claim 7, characterized in that: The system also includes a data acquisition and preprocessing module, which is used to collect the voltage and current of the electrical equipment, clean the voltage and current, and convert the cleaned voltage and current into active power and reactive power.

9. A non-intrusive renewable energy load forecasting method based on dual convolutional neural networks, characterized in that: It includes the following steps: Determine the state change event of the power consumption device through the sliding window event detection algorithm, and calculate the power difference before and after the event according to the state change event of the power consumption device, and use the power difference before and after the event as the load feature; Using a state convolutional neural network to extract the time series features and the spatial features in the load features, generating a predicted power state distribution according to the time series features and the spatial features, and inputting the predicted power state distribution into the power convolutional neural network to calculate the power consumption under the predicted power state distribution; The predicted power state distribution and power consumption are converted into the working state of each electrical device at each sampling time point, and a monitoring report is generated based on the working state.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to claim 9 are implemented.

Images