Real-time performance evaluation method and system for central air conditioner demand response equipment, medium and processor

Abstract

The invention discloses a real-time performance evaluation method and system for central air conditioner demand response equipment, a medium and a processor, and relates to the technical field of air conditioner performance evaluation. According to the method, equipment operation data are collected in real time and normalized, prediction data are generated by using a Transform model after time sequence segmentation, and real-time performance levels and probabilities are obtained through a one-dimensional convolutional neural network and an LSTM model after inverse normalization. The system comprises an acquisition module, a segmentation module, a prediction module, a restoration module, a capture module and an output module. A computer readable storage medium and a processor may execute the method. According to the method, the real-time performance of the equipment is dynamically evaluated, the resource allocation of the power system is optimized, stable operation is assisted, and the method has real-time performance and high efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of air conditioning performance evaluation, in particular to a real-time performance evaluation method, system, medium and processor of a central air conditioning demand response device. BACKGROUND

[0002] As a key link in the power industry, demand side management of power system aims to efficiently manage and optimize the load of terminal devices through intelligent means, to ensure the stable operation of the power system and optimize resource allocation. The peak load period is undoubtedly one of the biggest challenges faced by the power system operation, and it must quickly and accurately respond to load fluctuations to ensure the stability and reliability of power supply. In this context, it is particularly important to carry out real-time performance evaluation of demand response devices at the load terminal. Central air conditioning, as an important terminal device for power consumption, has become a major participant in demand response. Its large number and strong adjustable capacity effectively regulate power consumption, providing key support for load management and power dispatching during peak periods. Therefore, the real-time performance evaluation of central air conditioning demand response devices has become an important focus.

[0003] However, due to the complexity of power system operation and the continuous change of load demand, especially in the case of continuous growth of power demand, the load fluctuation and change of power grid are becoming increasingly complex, which is particularly evident during peak load periods. Traditional static performance evaluation methods cannot track and adapt to these rapid changes in real time, so they cannot meet the demand for real-time performance evaluation of the power system. In order to overcome this limitation, there is an urgent need for a method that can effectively evaluate the real-time performance of demand response devices.

[0004] In view of this, a real-time performance evaluation method, system, medium and processor of a central air conditioning demand response device are needed. SUMMARY

[0005] To solve the problem that the existing technology cannot meet the demand for real-time performance evaluation of the power system, the present application provides a real-time performance evaluation method, system, medium and processor of a central air conditioning demand response device, which can better support the dispatching work of the power grid, optimize the resource allocation of the power system, and help the stable operation of the power system. The specific technical solutions are as follows:

[0006] A real-time performance evaluation method of a central air conditioning demand response device, comprising:

[0007] S1: Real-time collection and normalization processing of the operation data of the central air conditioning device to obtain a normalized time series data set;

[0008] S2: Using a time series segmentation method to process the time series data set to obtain a segmented time series data segment set;

[0009] S3: positionally encode the time series data segments in the time series data segment set, and generate normalized prediction data Y of the future time period through the multi-head attention mechanism and the feedforward network of the Transformer model;

[0010] S4: perform inverse normalization processing on the normalized prediction data Y to restore it to the original physical quantity scale, to obtain prediction data of the original scale;

[0011] S5: perform convolution operation on the prediction data of the original scale through a one-dimensional convolutional neural network, to capture local change patterns in each time period, and obtain a local feature sequence;

[0012] S6: input the local feature sequence into an LSTM model for processing, to obtain a real-time performance level of the central air conditioning equipment and a corresponding probability.

[0013] Further, in step S2, the processing of the time series data set using the time series segmentation method refers to using a sliding window method to non-overlappingly segment the time series data in the time series data set at fixed time intervals, to form a plurality of non-overlapping time period data, and then to aggregate and average the time period data to reduce data fluctuations, to obtain a segmented time series data segment set.

[0014] Further, in step S3, the positionally encoding the time series data segments in the time series data segment set, and generating normalized prediction data Y of the future time period through the multi-head attention mechanism and the feedforward network of the Transformer model, includes the following steps:

[0015] S31: generate sinusoidal / cosine position encoding for each time point in the segmented data segment, and fuse the position encoding with the feature data, to obtain fused time series data X of each time point in the segmented data segment;

[0016] S32: input the fused time series data X into the Transformer model, to capture the dependency relationship of features at different time points in the sequence through the multi-head attention mechanism of the Transformer model, to generate context-aware feature representation;

[0017] S33: perform nonlinear transformation on the feature representation output by the attention mechanism, to capture local feature dependency, to obtain feature representation of the current segment;

[0018] S34: generate normalized prediction data of the future time period based on the feature representation of the current segment.

[0019] Further, in step S31, the formula of the sinusoidal / cosine position encoding is as follows:

[0020]

[0021] where PE refers to position encoding; pos is the position index of the current time point in the segment; i is the feature dimension index; d model is the total number of input feature dimensions.

[0022] Further, in step S32, the fusion time series data X is input into the Transformer model, and the multi-head attention mechanism of the Transformer model is used to capture the dependency relationship between different time points in the sequence, and generate context-aware feature representation, including the following steps:

[0023] S321: The fusion time series data X is respectively input into three linear layers to generate Query, Key and Value vectors;

[0024] S322: The Query, Key and Value vectors are divided into h independent heads, and the attention score Attention(Head i ) of each head is calculated separately:

[0025]

[0026] where Q i , K i , V i are the Query, Key and Value matrices of the i-th head, is a scaling factor;

[0027] S323: The attention scores of the h heads are concatenated and input into a linear layer to obtain the multi-head attention output Z attn :

[0028] Z attn = Concat(Head1,…,Head h )·W O ;

[0029] S324: The fusion time series data X and the multi-head attention output Z attn are added and then subjected to layer normalization to obtain the feature representation Z norm1 :

[0030] Z norm1 = LayerNorm(X+Z attn );

[0031] In the above formula, W O is a parameter matrix; Head i is the i-th attention score.

[0032] Further, in step S33, the feature representation of the attention mechanism output is nonlinearly transformed to capture local feature dependencies, obtaining the feature representation of the current segment, including the following steps:

[0033] S331: The feature dimension of the feature representation is raised to d by a linear layer and applying a ReLU activation function ff , as follows

[0034] Z ff1 = ReLU(Z norm1 ·W1+b1);

[0035] S332: Then, the dimension is reduced back to d by a linear layer model , as follows

[0036] Z ff2 = Z ff1 ·W2+b2;

[0037] S333: Again, residual connection and layer normalization are introduced to obtain the final feature representation Z final of the current segment, as follows

[0038] Z final = LayerNorm(Z norm1 +Z ff2 ).

[0039] Further, in step S34, the normalized prediction data of the future time period is generated based on the feature representation of the current segment, as follows

[0040]

[0041] In the above formula, is the normalized prediction data; is a parameter matrix.

[0042] A real-time performance evaluation system of a central air conditioning demand response device is applied to the real-time performance evaluation method of the central air conditioning demand response device, and includes:

[0043] A collection module is configured to collect and normalize the operation data of the central air conditioning device in real time, to obtain a normalized time series data set;

[0044] A segmentation module is configured to process the time series data set using a time series segmentation method, to obtain a segmented time series data segment set;

[0045] A prediction module is configured to positionally encode the time series data segments in the time series data segment set, and generate normalized prediction data Y of a future time period through a multi-head attention mechanism and a feedforward network of a Transformer model.

[0046] a reducing module configured to inverse normalize the normalized prediction data Y to restore the original physical quantity scale, to obtain prediction data in the original scale;

[0047] a capturing module configured to perform convolution operation on the prediction data in the original scale by a one-dimensional convolutional neural network, to capture local change patterns in each time period, and to obtain a local feature sequence;

[0048] an output module configured to input the local feature sequence into an LSTM model for processing, to obtain a real-time performance grade of the central air conditioning equipment and a corresponding probability.

[0049] A computer-readable storage medium, characterized in that the computer-readable storage medium comprises a stored program, wherein the program, when executed, controls the device in which the computer-readable storage medium is located to perform the real-time performance evaluation method of the central air conditioning demand response equipment described above.

[0050] A processor configured to execute a program, wherein the program, when executed, performs the real-time performance evaluation method of the central air conditioning demand response equipment described above.

[0051] Compared with the prior art, the beneficial effects of the present application are:

[0052] The present application provides a real-time and dynamic evaluation method, which can better support the dispatching work of the power grid, optimize the resource allocation of the power system, and help the stable operation of the power system. Specifically, this method can process and predict important time series data of central air conditioning, further classify and evaluate the data through a deep learning model, and realize the evaluation of the real-time performance of the equipment. This not only improves the operation efficiency of the power system, but also helps to realize the green and intelligent operation of the power system.

[0053] 2. Real-time dynamic evaluation capability: by collecting real-time operation data and combining Transformer, one-dimensional convolutional neural network (1D-CNN) and LSTM model, the dynamic tracking of equipment performance is realized, and the problem that traditional static evaluation methods cannot adapt to the rapid change of power system load is solved.

[0054] 3. Multi-model synergy advantage: the multi-head attention mechanism of Transformer captures long-term sequential dependence, 1D-CNN extracts local features, and LSTM models time trends. The combination of multiple models improves the processing accuracy of complex time series data and can accurately identify key patterns such as power fluctuations and temperature and humidity mutations.

[0055] 4. Optimize resource allocation and system stability: By predicting equipment performance levels, provide real-time data support for power grid dispatch, assist in peak load management, improve the efficiency of power system resource allocation, enhance operational stability, and promote the construction of a green and intelligent power system. Attached Figure Description

[0056] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0057] Figure 1 A flowchart illustrating a real-time performance evaluation method for a central air conditioning demand response device;

[0058] Figure 2 This is a schematic diagram of a real-time performance evaluation system for a central air conditioning demand response device. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] It should be understood that, when used in this application, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof.

[0061] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0062] It should also be further understood that the term “and / or” as used in this application refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes such combinations.

[0063] Example 1

[0064] like Figure 1 As shown, a real-time performance evaluation method for a central air conditioning demand response device includes the following steps:

[0065] S1: Real-time collection and normalization processing of operation data of the central air conditioning equipment to obtain a normalized time series data set.

[0066] Further, the operation data of the central air conditioning equipment includes multi-dimensional original data such as indoor temperature and humidity, outdoor temperature and humidity, total power consumption, equipment operating state, compressor state, etc.

[0067] Further, the normalization processing refers to using the minimum-maximum normalization method to map multi-dimensional original data of different dimensions into the same numerical range to obtain the normalized time series data set, and the normalization processing formula is as follows:

[0068]

[0069] In the above formula, D is the original data, D min and D max are the minimum and maximum values of the original data; and D * is the normalized data.

[0070] [Time point 1: (indoor temperature and humidity, outdoor temperature and humidity, power, equipment state, compressor state), time point 2:..., time point N:...]. Each time point contains 6-dimensional features (corresponding to the file "indoor temperature and humidity, outdoor temperature and humidity, total power consumption, equipment operating state, compressor state").

[0071] Assuming that the original time series data contains 12 time points (numbered 0-11), the normalized values are simplified as follows (only the first 3-dimensional features are shown, and the equipment state and compressor state are Boolean values):

[0072] Time point Indoor temperature (normalized) Indoor humidity (normalized) Power (normalized) Device status Compressor status 0 0.6 0.4 0.8 1 1 1 0.7 0.5 0.9 1 1 2 0.5 0.3 0.7 1 0 3 0.4 0.2 0.6 0 0 4 0.3 0.1 0.5 0 0 5 0.2 0.3 0.4 0 1 6 0.1 0.4 0.3 1 1 7 0.2 0.5 0.4 1 1 8 0.3 0.6 0.5 1 0 9 0.4 0.7 0.6 0 0 10 0.5 0.8 0.7 0 0 11 0.6 0.9 0.8 0 1

[0073] S2: Using a time series segmentation method to process the time series data set to obtain a segmented time series data segment set.

[0074] Further, the processing of the time series data set using the time series segmentation method refers to using a sliding window method to non-overlappingly cut the time series data in the time series data set at a fixed time interval (such as 15 minutes) to form a plurality of non-overlapping time period data, and then aggregating and averaging each time period data to reduce data fluctuations to obtain the segmented time series data segment set.

[0075] Sliding window parameters:

[0076] Window size: set to N = 4 time points (such as each window corresponds to 15 minutes x 4 = 1 hour).

[0077] Step size: Set to t = 4 time points (no overlap split, i.e. the next window starts after the current window ends).

[0078] Prediction scale: Set to l = 2 time periods (i.e. predict the trend of the next 2 windows).

[0079] Segmentation logic: For time series data of length T, split into segments at fixed intervals, each containing 6-dimensional features of consecutive N time points. For example, 12 time points in step S1 are segmented into 3 time series data segments with window size = 4, step size = 4, as follows:

[0080] Segment 1:

[0081]

[0082] Segment 2:

[0083]

[0084] Segment 3:

[0085]

[0086] S3: Encode the position of the time series data segments in the time series data segment set, and generate normalized prediction data for the future time period through the multi-head attention mechanism and feed-forward network of the Transformer model.

[0087] The Transformer model is composed of a multi-head attention mechanism (Multi-Head Attention) and a feed-forward neural network (Feed-Forward Network, FFN).

[0088] S31: Generate sine / cosine position encoding for each time point in the segmented data segment, and fuse the position encoding with the feature data to obtain the fused time series data for each time point in the segmented data segment. The position encoding calculation formula is:

[0089]

[0090] wherein PE (Position Embedding) refers to position encoding, which is used to encode the time (pos represents the time point) and position (i represents the feature position) information in the time series data into numerical features recognizable by the model, so that the Transformer model can capture the time sequence order of the data; pos is the position index of the current time point in the segment (starting from 0, for example, the first time point in the segment pos = 0, the second pos = 1); i is the feature dimension index (corresponding to the 6-dimensional feature in step S2, for example, i = 0 is the indoor temperature, i = 1 is the indoor humidity, i = 2 is the power, i = 3 is the device status, i = 4 is the compressor status, and i = 5 is the reserved expansion dimension).d. model is the total dimension number of input features, which is 6 in the cited document. 2i indicates that the sine function is used for even numbers, and 2i+1 indicates that the cosine function is used for odd dimensions, that is:

[0091] The indoor temperature (i = 0) corresponds to the encoding dimensions 2i = 0 (sine) and 2i+1 = 1 (cosine), and there are two corresponding encoding dimensions (even and odd subscripts) for the feature temperature.

[0092] The indoor humidity (i = 1) corresponds to the encoding dimensions 2i = 2 (sine) and 2i+1 = 3 (cosine), and there are two corresponding encoding dimensions (even and odd subscripts) for the feature humidity.

[0093] Similarly, until i = 5 (such as the compressor status) corresponds to the encoding dimensions 10 (sine) and 11 (cosine), and there are two corresponding encoding dimensions (even and odd subscripts) for the feature compressor status. Each feature i (i = 0, 1, 2, 3, 4, 5) has two corresponding encoding dimensions (even and odd subscripts), so the total encoding dimension is 2xd model = 12.

[0094] The following is an example of partial position encoding (taking segment 1 of step S2 as an example, only one encoding dimension for each feature dimension is listed):

[0095] Segment 1 contains 4 time points (pos = 0, 1, 2, 3) and 6-dimensional features (i = 0-5).

[0096] Calculate the position encoding of the indoor temperature (i = 0):

[0097] When pos = 0, PE (0,0) = sin(0 / 10000 0 / 6 ) = sin(0) = 0;

[0098] When pos = 1, PE (1,0) = sin(1 / 10000 0 / 6 ) = sin(0.0001) ≈ 0.0001;

[0099] The position encoding of indoor humidity (i = 1) is calculated as:

[0100] When pos = 0, PE (0,1) = cos(0 / 10000 2×1 / 6 ) = cos(0) = 1.

[0101] When pos = 1, PE (1,1) = cos((1 / 10000 2×1 / 6 ) = cos((1 / 10000 1 / 3 ) ≈ cos(0.1) ≈ 0.995.

[0102] Further, the fusion of the position encoding and the feature data refers to summing the original feature value of each feature i at each time point with its position encoding value, and the resulting summation value is taken as the new feature value of the feature i, thereby obtaining the fusion time series data of each time point in the segmented data segment, denoted as N is the number of time points within the segment, d model is the feature dimension after fusion. For example, the indoor temperature feature at the first time point (pos = 0) in segment 1 is: Indoor temperature feature = normalized value (0.6) + PE (0,0) = 0.6 + 0 = 0.6.

[0103] X = [x0+PE (pos,0) , x1+PE (pos,1) , …, x5+PE ((pos,10) , x5+PE (pos,11) ].

[0104] Wherein: x0-x5 are the 6-dimensional original features (such as indoor temperature, humidity, etc.) normalized in step S1; PE ((pos,0) , PE (pos,1) are the sine and cosine encodings corresponding to the indoor temperature (i = 0, subscript 2i = 0, 2i+1 = 1); PE ((pos,10) , PE ((pos,11) are the sine and cosine encodings corresponding to the compressor state (i = 5, subscript 2i = 10, 2i+1 = 11).

[0105] S32: input the fusion time series data into the Transformer model, and capture the dependency relationship of features at different time points in the sequence through the multi-head attention mechanism of the Transformer model to generate context-aware feature representation.

[0106] S321: respectively pass the fusion time series data X through three linear layers (parameter matrices W Q , W K , W V), generate Query (query vector), Key (key vector), Value (value vector), shape is (d k is the dimension of a single attention head.

[0107] S322: divide Query, Key, Value into h independent heads (take h = 2 in the example), and calculate the attention score of each head separately:

[0108]

[0109] Where, Q i , K i , V i are the Query, Key, Value matrices of the i-th head, is a scaling factor to avoid gradient vanishing.

[0110] S323: concatenate the attention scores of the h heads and pass them through a linear layer (parameter matrix W O ) to get the multi-head attention output

[0111] Z attn = Concat (Head1, …, Head h ) · W O ;

[0112] S324: To avoid gradient vanishing, add the fused time series data X and the multi-head attention output Z attn , and then perform layer normalization (LayerNorm) to get the feature representation:

[0113] Z norm1 = LayerNorm (X + Z attn );

[0114] S33: Perform nonlinear transformation on the feature representation output by the attention mechanism to capture local feature dependencies and get the feature representation of the current segment.

[0115] S331: Increase the feature dimension of the feature representation to d ff (d ff is the dimension of the feedforward layer, and d ff = 256 in the example) by applying a linear layer and ReLU activation function:

[0116] Z ff1 = ReLU (Z norm1 · W1 + b1);

[0117] S332: Then reduce the dimension to d model dimension by a linear layer:

[0118] Z ff2 = Z ff1 · W2+ b2;

[0119] S333: Introduce the residual connection and layer normalization again to get the final current segment feature representation Z final :

[0120] Z final = LayerNorm(Z norm1 + Z ff2 );

[0121] S34: Based on the current segment feature representation generate normalized prediction data of future l time periods (l = 2 in the example)

[0122] Specifically, the following steps are included: in the decoder part of the Transformer model (or through an independent linear layer), the feature representation Z final of the current segment is mapped to the feature space of the future time points to obtain the prediction value.

[0123] Further, by connecting a linear layer (parameter matrix W pred ) after Z final , the normalized prediction data of the future l time points

[0124]

[0125] wherein each row of corresponds to a 6-dimensional feature prediction value (such as indoor temperature, power, etc.) of a future time point.

[0126] S4: Perform inverse normalization processing on the normalized prediction data to restore it to the original physical quantity scale to obtain prediction data in the original scale. The normalized prediction data (output by the Transformer model, ranging from [0, 1]) is restored to the original physical quantity scale (such as °C for temperature, kW for power, etc.) so that subsequent local feature extraction and performance evaluation can be based on the actual dimension.

[0127] S41: Inverse normalization formula and principle.

[0128] According to the min-max normalization method adopted in step S1, the inverse process (inverse normalization) formula is:

[0129]

[0130] Parameter description, For normalized prediction data (model output values); D min D max The minimum and maximum values ​​of the original data recorded in step S1 (stored independently by feature dimension); D pred This is the original scale of the prediction data.

[0131] S42: Logic for Independent Restoration of Multi-Dimensional Data

[0132] Since step S1 in the document normalizes the six-dimensional features, including indoor temperature and humidity, outdoor temperature and humidity, total power consumption, equipment operating status, and compressor status, the inverse normalization needs to be processed column by column according to the feature dimensions. The specific steps are as follows:

[0133] 1. Store the original feature statistics: Before normalization in step S1, record D separately for each feature dimension. min and D max .

[0134] example:

[0135] Total power consumption original range: [10kW, 100kW], corresponding to D min =10,D max =100.

[0136] Original indoor temperature range: [18°C, 28°C], corresponding to D min =18,D max =28.

[0137] 2. Dimensional inverse normalization: For the predicted data (l represents the number of prediction time periods, d) model For each column of (=6), use the D of its corresponding dimension. min and D max Perform the restoration.

[0138] like If the i-th column corresponds to the normalized predicted value of "total power consumption", then the restoration formula is:

[0139]

[0140] If column j corresponds to "indoor temperature", then:

[0141]

[0142] S43: Boolean Feature Processing Rules

[0143] In the file, "Equipment Operating Status" and "Compressor Status" are Boolean features (0 or 1 after normalization). During inverse normalization, they are directly converted to the original state description.

[0144] 0 → Close

[0145] 1→ running

[0146] Example: If the "compressor status" normalized value at a certain time point in the prediction data is 0.98 (close to 1), it is determined as "running" after inverse normalization.

[0147] S44: Outlier processing mechanism

[0148] If the model's predicted normalized value exceeds the range (0, 1) (such as due to training errors), it needs to be clipped to the interval boundary before inverse normalization: if set to 0; if set to 1 to avoid physically unreasonable results after restoration (such as negative power).

[0149] S5: Perform convolution operation on the original scale of the prediction data through one-dimensional convolutional neural network (1D-CNN), capture local change patterns (such as power fluctuation peaks, temperature and humidity mutations) in each time period, and obtain local feature sequence.

[0150] 1. Convolution layer design

[0151] Convolution kernel configuration:

[0152] Number: Set 32 convolution kernels to capture different local patterns in multi-dimensional features.

[0153] Size: Choose 3x1 convolution kernel (slide along time dimension, process 3 consecutive time points of the same feature each time), ensure to capture short-term dependence of adjacent time points (such as power fluctuation within 15 minutes).

[0154] Step: Set step to 1 to ensure no overlapping extraction of features and preserve the integrity of time series.

[0155] Activation function: Use ReLU activation function to enhance the model's nonlinear expression ability, formula:

[0156] f(x) = max(0, x)

[0157] Set to "same" mode to keep the output feature map consistent with the input dimension, avoiding edge information loss.

[0158] 2. Feature extraction process

[0159] Multi-feature parallel convolution:

[0160] Convolution operation is performed on 6-dimensional features (such as power, temperature, etc.), and each convolution kernel generates an independent feature map (Feature Map).

[0161] Example: If the input is 6-dimensional data for 2 time periods (shape 2x6), after processing by 32 3x1 convolution kernels, 32 feature maps are output, with a shape of 2x32.

[0162] Local pattern capture:

[0163] Power feature: Capture power peaks or abnormal fluctuations in a short period of time (such as a sudden increase in power when the compressor starts).

[0164] Temperature and humidity feature: Identify temperature and humidity mutations (such as a sudden increase in outdoor temperature leading to changes in indoor air conditioning load).

[0165] Device status feature: Combine state changes at consecutive time points (such as device start-stop frequency) to judge running stability.

[0166] 3. Pooling layer optimization (optional)

[0167] Max pooling: Add a pooling layer after the convolution layer (pooling window size 2), reduce feature dimension, reduce calculation amount, formula:

[0168] Pooled(x) = max(x i ,x i+1 );

[0169] Pooled(x) represents the output feature value after pooling, i.e. the maximum value of the feature values at adjacent time points; x i , x i+1 represent the feature values at adjacent time points in the feature map (continuous elements along the time dimension).

[0170] Effect: Retain the maximum value (such as peak power) in local features, suppress noise, and highlight key change points.

[0171] 4. The local feature map output by the convolution layer needs to be flattened into a one-dimensional vector as input for the LSTM model.

[0172] Example: If the convolution output is a 2x32 feature map, flattening it results in a 64-dimensional vector, corresponding to a 32-dimensional local feature sequence for each time period.

[0173] S6: Input the local feature sequence into the LSTM model for processing to obtain the real-time performance rating of the central air conditioning equipment and the corresponding probability.

[0174] After completing the prediction, the performance of the central air conditioning demand response equipment is comprehensively classified and evaluated. By inputting the predicted data into the one-dimensional convolution layer, LSTM layer, and fully connected layer, the system can perform multi-level processing and analysis on the performance of the equipment. Finally, the model will evaluate the performance of the equipment based on real-time data and output the corresponding classification results.

[0175] S61: Time-dependent modeling of local feature sequences using LSTM layers to extract overall trend features and long-term relationships of the data (e.g., daily energy consumption patterns, equipment operation stability).

[0176] S62: Map the overall trend features output by LSTM to the pre-set performance levels (e.g., "excellent", "good", "average", "poor"), and generate the probability distribution of each level through the Softmax activation function, i.e., obtain the real-time performance level of the central air conditioning equipment and the corresponding probability.

[0177] The time-dependent modeling and classification module is based on one-dimensional convolution to extract local features, perform time-dependent modeling, capture the overall trend and long-term relationship of the data, and the fully connected neural network receives the feature information output by LSTM and performs classification through the Softmax activation function, outputting multiple levels of probabilities of performance evaluation.

[0178] The performance level is divided into four levels: "excellent", "good", "average", and "poor", which reflect the actual performance of the equipment in demand response. LSTM layer captures time series features, combined with the classification function of the fully connected layer, the model can automatically classify the performance of the equipment into the corresponding level according to the real-time data. To quantify the effectiveness of this evaluation mechanism, we introduce the F1 score, precision, and recall indicators to evaluate the performance classification effect of the model, as shown in Table 1.

[0179] Table 1: Indicator Evaluation Table

[0180] F1 Score Precision Recall 94.7％ 95.7％ 93.6％

[0181] Further to be explained:

[0182] GRU is a simplified version of LSTM, which improves processing efficiency by reducing calculation steps. It has a gating mechanism similar to LSTM, which can capture short-term dependencies, but its memory capacity is weaker, especially when dealing with complex, long-term dependent time series data.

[0183] LSTM layer: Although the calculation efficiency of LSTM is not as good as GRU, its ability to capture long-term dependencies is stronger, and it performs more stably in complex time series data and classification tasks. LSTM can remember longer time series information, reducing information loss in classification tasks, so for tasks such as central air conditioning demand response with strong long-term dependence, the classification performance of LSTM is better than that of GRU.

[0184] GRU layer: GRU has higher computational efficiency, and its simplified structure makes it effective in handling short-term dependent time series data. However, GRU is not as good as LSTM in capturing long-term dependencies, especially in applications involving complex temporal patterns and long time spans, GRU may miss some key information, affecting the accuracy of classification.

[0185] Therefore, in the classification task, although LSTM is slightly slower in efficiency, its accuracy in handling long-term dependencies is more important, especially in the performance evaluation of central air conditioning demand response devices, the long-term dependence of time series data is strong. LSTM can effectively capture these long-term dependencies, thereby improving classification performance. While GRU is more efficient, its classification accuracy may not be sufficient when dealing with complex, long time series, so in this scenario, the combination of one-dimensional convolution layer + LSTM layer + fully connected layer is still a better choice.

[0186] The present application can process the running data of the central air conditioner at a fixed time interval by introducing a time series segmentation method, so that the data fluctuation in each time period is smoothed, thereby improving the system's ability to capture short-term trends. Combined with the multi-head attention mechanism of the Transformer model, the indoor and outdoor temperature, humidity and power data are analyzed globally, improving the accuracy and computational efficiency of data prediction, and adapting to the demand for large-scale data parallel processing. The present application uses inverse normalization operation to restore the normalized data output by the model to the actual physical quantity, ensuring that the analysis results can match the actual state. By using a one-dimensional convolution layer to extract local features, the data change pattern in each time period is captured, further optimizing the local processing effect of the data. Finally, the LSTM layer models the long-term time dependence, combines local features into global trends, and classifies through the fully connected layer. The system can quickly output multiple performance level classification results, helping users understand the running state and energy efficiency of the air conditioner in a timely manner, and improving the response speed to grid dispatching demand.

[0187] Embodiment two

[0188] As Figure 2 A real-time performance evaluation system for a central air conditioning demand response device, applied to the real-time performance evaluation method of the central air conditioning demand response device described above, comprising:

[0189] The acquisition module is used for real-time acquisition and normalization processing of the running data of the central air conditioning device, and obtains a normalized time series data set;

[0190] The segmentation module is used for processing the time series data set using a time series segmentation method to obtain a segmented time series data segment set;

[0191] a prediction module configured to positionally encode time series data segments in a set of time series data segments, and generate normalized prediction data Y of a future time period through a multi-head attention mechanism and a feedforward network of a Transformer model;

[0192] a reduction module configured to inversely normalize the normalized prediction data Y to restore the original physical quantity scale, to obtain prediction data of the original scale;

[0193] a capture module configured to perform convolution operation on the prediction data of the original scale through a one-dimensional convolutional neural network, to capture local change patterns in each time period, and to obtain a local feature sequence;

[0194] an output module configured to input the local feature sequence into an LSTM model for processing, to obtain a real-time performance grade of the central air conditioning equipment and a corresponding probability.

[0195] Embodiment Three

[0196] A computer-readable storage medium, characterized in that the computer-readable storage medium comprises a stored program, wherein the program controls the device where the computer-readable storage medium is located to execute the real-time performance evaluation method of the central air conditioning demand response equipment described above when the program is running.

[0197] Embodiment Four

[0198] A processor configured to run a program, wherein the program executes the real-time performance evaluation method of the central air conditioning demand response equipment described above when the program is running.

[0199] The present application discloses a real-time performance evaluation method, system, medium and processor of a central air conditioning demand response equipment, which relates to the technical field of air conditioner performance evaluation. The method generates prediction data through a Transformer model after real-time collection of equipment operation data and normalization and time series segmentation, and obtains a real-time performance grade and a probability through a one-dimensional convolutional neural network and an LSTM model after inverse normalization. The system comprises acquisition, segmentation, prediction, reduction, capture and output modules. The computer-readable storage medium and the processor can execute the method. The present application realizes dynamic evaluation of real-time performance of the equipment, optimizes resource allocation of the power system, helps stable operation, and has real-time performance and high efficiency.

[0200] Those skilled in the art can understand that the units of the examples described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components of the examples have been described in general terms in the above description. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0201] In the embodiments provided by the present application, it should be understood that the division of units is only a logical functional division. In actual implementation, there can be another division manner. For example, multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored, etc.

[0202] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0203] When the integrated unit is realized in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0204] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and they should be covered in the scope of the present application.

Claims

1. A method for evaluating real-time performance of a central air conditioning demand response device, characterized in that, The method comprises the following steps: S1: real-time collection and normalization processing of operation data of a central air conditioning device to obtain a normalized time series data set; S2: processing the time series data set using a time series segmentation method to obtain a segmented time series data segment set; S3: position encoding of the time series data segments in the time series data segment set, and generation of normalized prediction data Y of a future time period through a multi-head attention mechanism and a feedforward network of a Transformer model; S4: inverse normalization processing of the normalized prediction data Y to restore it to an original physical quantity scale to obtain prediction data of the original scale; S5: convolution operation on the prediction data of the original scale through a one-dimensional convolutional neural network to capture local change patterns in each time period to obtain a local feature sequence; S6: inputting the local feature sequence into an LSTM model for processing to obtain a real-time performance level of the central air conditioning device and a corresponding probability.

2. The method of claim 1, wherein, In step S2, the processing of the time series data set using the time series segmentation method refers to using a sliding window method to non-overlappingly divide the time series data in the time series data set at a fixed time interval to form a plurality of non-overlapping time period data, and then aggregating and averaging the time period data to reduce data fluctuations to obtain the segmented time series data segment set.

3. The method of claim 1, wherein, In step S3, the position encoding of the time series data segments in the time series data segment set and the generation of the normalized prediction data Y of the future time period through the multi-head attention mechanism and the feedforward network of the Transformer model comprise the following steps: S31: generating sinusoidal / cosine position encoding for each time point in the segmented data segment, and fusing the position encoding with the feature data to obtain fusion time series data X for each time point in the segmented data segment; S32: inputting the fusion time series data X into the Transformer model to capture the dependency relationship of features at different time points in the sequence through the multi-head attention mechanism of the Transformer model to generate context-aware feature representation; S33: performing nonlinear transformation on the feature representation output by the attention mechanism to capture local feature dependency to obtain feature representation of the current segment; S34: generating normalized prediction data of a future time period based on the feature representation of the current segment.

4. The method of claim 3, wherein, In step S31, the formula of the sinusoidal / cosine position encoding is as follows: wherein PE refers to position encoding; pos is the position index of the current time point within the segment; i is the feature dimension index; d model is the total number of input feature dimensions.

5. The method of claim 3, wherein, In step S32, the inputting of the fusion time series data X into the Transformer model to capture the dependency relationship of features at different time points in the sequence through the multi-head attention mechanism of the Transformer model to generate context-aware feature representation comprises the following steps: S321: generating Query, Key, and Value vectors through three linear layers on the fusion time series data X, respectively; S322: divide the Query, Key, Value vectors into h independent heads, each head separately computes attention score Attention(Head i ): where Q i , K i , V i are the Query, Key, Value matrices for the i-th head, is a scaling factor; S323: Concatenate the attention scores of h heads and pass through a linear layer to get the multi-head attention output Z attn : Z attn = Concat(Head1,..., Head h ) · W O ; S324: fuse the timing data X with the multi-head attention output Z attn layer normalization to obtain the feature representation Z norm1 : Z norm1 = LayerNorm(X + Z attn ); In the above formula, W O is a parameter matrix; Head i is the i-th head attention score.

6. The method of claim 5, wherein, In step S33, the nonlinear transformation on the feature representation output by the attention mechanism to capture local feature dependency to obtain the feature representation of the current segment, In step S34, the formula for generating normalized prediction data of a future time period based on the feature representation of the current segment is as follows: S331: Boost the feature dimension of the feature representation to d by a linear layer and applying a ReLU activation function ff , as follows Z ff1 = ReLU(Z norm1 · W1+ b1); S332: Then reduce back to d through a linear layer model dimension, as follows: Z ff2 = Z ff1 • W2 + b2; S333: Introduce the residual connection and layer normalization again to get the final current segment feature representation Z final , as follows Z final = LayerNorm(Z norm1 + Z ff2 ).

7. The method of claim 6, wherein the method further comprises: ​ In the above formulae, is the normalized prediction data; is the parameter matrix.

8. A real-time performance evaluation system for a central air conditioning demand response device, characterized by, The real-time performance evaluation method applied to the central air conditioning demand response device of any one of claims 1 to 7 comprises: a collection module for collecting and normalizing the operation data of the central air conditioning device in real time to obtain a normalized time series data set; a segmentation module for processing the time series data set using a time series segmentation method to obtain a segmented time series data segment set; a prediction module for positionally encoding time series data segments in a set of time series data segments and generating normalized prediction data for a future time period through a multi-headed attention mechanism of a Transformer model and a feed-forward network a reduction module for reducing the normalized predicted data to the original physical quantity scale by inverse normalization to obtain predicted data in the original scale; a capture module for performing convolution operation on the original scale prediction data by a one-dimensional convolutional neural network to capture the local change pattern in each time period and obtain a local feature sequence; an output module for inputting the local feature sequence into an LSTM model for processing to obtain a real-time performance grade of the central air conditioning device and a corresponding probability.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium comprises a stored program, wherein when the program runs, the device where the computer readable storage medium is located executes the real-time performance evaluation method of the central air conditioning demand response device of any one of claims 1 to 7.

10. A processor, comprising: The processor is used to run a program, wherein when the program runs, the real-time performance evaluation method of the central air conditioning demand response device of any one of claims 1 to 7 is executed.

Images