Low-dew-point double-rotating-wheel dehumidifier energy-saving prediction method based on time-space diagram attention

Abstract

The invention discloses a low-dew-point double-rotating-wheel dehumidifier energy-saving prediction method based on time-space diagram attention. The invention provides a time-space diagram attention neural network based on an attention mechanism. According to the time-space diagram attention neural network based on the attention mechanism, a diagram attention network, a bidirectional gating circulation unit and a time sequence Transform model are combined. The method comprises the following steps: firstly, constructing a graph structure according to equipment topological structure information, vectorizing a time slice input matrix, introducing an attention mechanism, and adaptively distributing node weights to update node features; adopting two forward and reverse unidirectional GRUs structures, and fusing past information and future information for feature prediction; and finally, through position embedding and a time sequence Transform structure, hidden information between sliding window inputs is mined, loss of effective features is reduced, and the long-term prediction capability of the model is improved. According to the method, the spatial-temporal correlation of original data can be better captured, the equipment operation state parameters are effectively predicted, and the energy-saving optimization control process is guided.

Description

Technical Field

[0001] The present invention belongs to the technical field of dehumidification energy-saving control and is a method for optimizing the control of a low dew-point dual-rotor dehumidifier. Specifically, it relates to a method for optimizing the control of a low dew-point dual-rotor dehumidifier based on a spatiotemporal graph attention neural network. Background Art

[0002] The low dew point double-rotor dehumidifier is a special type of rotary dehumidification equipment composed of a double-rotor structure. As a dehumidification equipment for low humidity environments, it is widely used in industrial production, warehousing and other related occasions with strict requirements on environmental humidity. In some production process links, the environmental relative humidity of the product is required to be lower than 2%, and the environmental dew point temperature is required to be lower than -28°C. At present, in response to such low dew point environmental requirements, the existing technology generally adopts low dew point double-rotor dehumidification technology, including the following steps: 1) In the air treatment section, the fresh air is cooled and dehumidified by the front surface cooler to reduce the fresh air temperature, and then enters the first-stage rotor of the low dew point double-rotor dehumidifier for preliminary dehumidification, and is mixed with the room return air after reaching a certain quality. The mixed air enters the middle surface cooler for cooling and then enters the secondary wheel for deep dehumidification. Finally, it is cooled by the rear surface cooler or heated by the regeneration heater according to the set temperature requirements before entering the room; 2) In the regeneration air section, part of the air dehumidified by the secondary wheel enters the low dew point double wheel dehumidifier to become regeneration air. It is first heated by the rear regeneration heater and enters the regeneration area of ​​the secondary wheel to take away the moisture in the secondary wheel, restore the dehumidification performance of the secondary wheel, and is mixed with the regenerated fresh air, which enters the low dew point double wheel dehumidifier after dehumidification by the first wheel. The mixed air is then heated by the front regeneration heater and enters the regeneration area of ​​the first wheel to take away the moisture in the first wheel. Finally, it is discharged to the outside through the regeneration fan.

[0003] Currently, low dew-point dual-rotor dehumidifiers are complex systems with nonlinearity, large inertia, and strong coupling. However, the actual control process mostly adopts PID control method, which requires manual adjustment, leaves too large control margin, and lacks refined, intelligent, and dynamic control methods. As a result, strong time lag and high energy operating costs are generated during the actual operation of the system. Summary of the invention

[0004] In order to overcome the energy-saving optimization control problem of spatiotemporal air state data of dehumidification equipment existing in the prior art, the present invention provides a low dew point dual-rotor dehumidifier energy-saving optimization control prediction method based on spatiotemporal graph attention neural network, which predicts the air state of each data collection point of the current low dew point dual-rotor dehumidifier, and provides energy-saving optimization control guidance for the surface cooler, regeneration heater, etc. under the conditions of ensuring the temperature and humidity of the control object, so as to achieve the energy-saving optimization control goal.

[0005] The technical solution of this technology is:

[0006] 1. A method for energy saving prediction of low dew point double-rotor dehumidifier based on spatiotemporal graph attention

[0007] S1: Sensors are placed at key positions of the low dew point dual-rotor dehumidifier to collect air status data and equipment control parameters of the surface cooling and regeneration heater at the above key positions in real time, and the dehumidifier historical operation status data set is constructed based on the air status data and equipment control parameters of the surface cooling and regeneration heater collected in real time by the sensors;

[0008] S2: Taking the layout positions of the low dew point dual-rotor dehumidifier sensors as nodes and the air circulation channels inside the rotary dehumidifier as edges, an undirected position relationship graph is constructed based on the relative relationship between the sensor layout positions and the air circulation channels.

[0009] S3: preprocess the historical operation status data set of the dehumidifier according to sampling time, node location, and data type to obtain a historical operation status database;

[0010] S4: Construct a neural network training set based on the historical operation status database;

[0011] S5: Construct a spatiotemporal graph attention neural network model based on an undirected graph structure, and use the neural network training set to train the spatiotemporal graph attention neural network model based on an undirected graph structure to obtain a trained equipment operation status prediction model;

[0012] S6: Generate a vectorized input matrix based on the real-time collected air state data and equipment control parameters, process the vectorized input matrix using a sliding window, and then input it into the trained equipment operation state prediction model. The output of the model includes the equipment operation state at the future moment and the optimized equipment control parameters of the surface cooling and regeneration heater.

[0013] In S1, the air state data includes air temperature, humidity, air volume or other collectible air state data, and the equipment control parameters are set temperature values ​​of the surface cooler and the regeneration heater.

[0014] The S3 is specifically:

[0015] The historical operation status data is cleaned, standardized, reconstructed and vectorized at each sampling moment to obtain a historical operation status database.

[0016] In S3, node feature reconstruction includes:

[0017] For the data after data standardization, the air state data at each moment is recorded as the first F-1 features of that moment. If the collection position is the air outlet position through the cooler or regeneration heater, the set temperature value of the cooler or regeneration heater is used as the Fth feature of that moment. Otherwise, the Fth feature of that moment is set to 0, thereby reconstructing the feature dimension of the collection position to Fth.

[0018] In S3, data vectorization includes:

[0019] For the data reconstructed by node features, the reconstructed single-moment operating status data is constructed into an operating status input matrix in sequence according to the time label; then, according to the node position label, the position relationship undirected graph and the operating status input matrix are fused to obtain a vectorized input matrix.

[0020] The process of processing the vectorized input matrix by using a sliding window includes:

[0021] The vectorized input matrices are respectively processed by sliding using a sliding window, thereby obtaining several input matrices of the same size.

[0022] In S5, the spatiotemporal graph attention neural network model based on an undirected graph structure includes a graph attention network, a bidirectional gated recurrent unit and a temporal Transformer network layer connected in sequence, and a multi-layer fully connected network for model output.

[0023] 2. A low dew point double-rotor dehumidifier energy-saving prediction system based on spatiotemporal graph attention

[0024] The operation status data collection module is used to collect the air status data and equipment control parameters at the key positions of the low dew point double-rotor dehumidifier in real time, so as to construct the operation status data set;

[0025] A position relationship undirected graph generation module is used to generate a position relationship undirected graph according to the topological relationship of the low dew point dual-rotor dehumidifier structure;

[0026] A data preprocessing module, used to convert the operating status data set into a vectorized input matrix;

[0027] The equipment operation status prediction module is used to store the equipment operation status prediction model and predict the status of the low dew point dual-rotor dehumidifier according to the input vectorized input matrix, so as to obtain the predicted equipment operation status at the future moment and the optimized equipment control parameters of the surface cooler and the regeneration heater.

[0028] 3. A computer device

[0029] The device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the energy-saving prediction method for a low dew point dual-rotor dehumidifier based on spatiotemporal graph attention are implemented.

[0030] 4. A computer-readable storage medium

[0031] The medium stores a computer program, which, when executed by a processor, implements the steps of a method for predicting energy conservation of a low dew-point dual-rotor dehumidifier based on spatiotemporal graph attention.

[0032] The beneficial effects of the present invention are as follows:

[0033] (1) In the spatial dimension, by extracting the topological structure information of the low dew point dual-rotor dehumidification equipment, embedding the time dimension, feature dimension and position encoding, introducing the masked self-attention mechanism and the multi-head attention mechanism, stacking the graph attention layer that focuses on the node neighborhood features, and using the neighborhood aggregation operation, it is allowed to assign different weights to different nodes in the neighborhood when processing different neighborhoods to update the central node features. This allows the model to efficiently capture the spatial correlation of the device nodes while being able to cope with scenarios where some node features are missing or some data are missing, greatly improving the model's expressive power. For situations where the device topology information is unknown or partially unknown, the model also has a certain expressive power.

[0034] (2) In the time dimension, by embedding two-layer GRUs, position encoding, multi-head attention mechanism, residual connection and layer normalization, the bidirectional gated recurrent unit BiGRU and the time series Transformer layer are used to effectively capture the local and global time correlation of the node time series data. The forward and reverse unidirectional GRUs are used respectively to effectively capture the influence of past information and future information on the current moment data, and to fully explore the information of the air state of the dual-rotor dehumidification equipment; the time series Transformer structure with multi-head attention mechanism is used to capture the global time correlation between the sliding window input data through position encoding, feature concatenation and attention fusion, which avoids the loss of effective node features to a certain extent and improves the prediction accuracy of the model.

[0035] (3) The model is reconstructed through node feature reconstruction, with controller setting parameters as input node features and air state parameters at each data collection point in the future as output. The predicted values ​​of some node controller setting parameters are used to assist in judging the changing trend of the air state at each point of the equipment and the direction of energy-saving optimization, thereby greatly improving the response speed of the dehumidification system and effectively guiding the energy-saving optimization control process.

[0036] Therefore, the present invention can be directly applied to low dew point dual-rotor dehumidification equipment, can capture the spatial correlation and temporal correlation of air state data, effectively predict the changing trend of air state, provide energy-saving optimization control strategy, and effectively reduce equipment energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 The present invention is a flow chart of a method for predicting low dew point dual-rotor dehumidification equipment.

[0038] Figure 2 Schematic diagram of the structure of a typical low dew point dual-rotor dehumidification device according to an embodiment of the present invention.

[0039] Figure 3 It is a schematic diagram of the position relationship undirected graph structure of an embodiment of the present invention.

[0040] Figure 4 Schematic diagram of the spatiotemporal graph attention neural network model based on undirected graph structure proposed in this invention. DETAILED DESCRIPTION

[0041] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0042] like Figure 1 As shown, the present invention comprises the following steps:

[0043] S1: Several sensors are arranged at key positions where the air conditions of the low dew point dual-rotor dehumidifier, such as the air inlet, air outlet, and mixed air, are significantly different, for real-time collection of air condition data. The sensors collect the air condition data at the above key positions in real time and record the control parameters of the current equipment surface cooler and regeneration heater. The dehumidifier historical operation status data set is constructed by collecting the historical air condition data and equipment control parameters; the air condition data collected by the sensors include but are not limited to air temperature, humidity, and air volume, and the control parameters of the surface cooler and regeneration heater in the current equipment should be recorded at the same frequency when recording data. The above key positions specifically include the controller, that is, the front and rear positions of the surface cooler / regeneration heater in the equipment, and the equipment fresh air inlet, room return air outlet, regeneration air supply outlet, fresh air rotor and low humidity cooling outlet, and equipment internal mixed air outlet. Equipment control parameters include but are not limited to the control temperature set for the surface cooler and regeneration heater. The air status and equipment parameter data of each node collected at a single time are taken as a single time slice, and the historical operation status data set is constructed by several time slices. Before the implementation of the plan, it is necessary to ensure that the number of collected data samples (i.e., the size of the historical status data set) is sufficient to support the training needs of the proposed deep learning fusion model.

[0044] S2: Based on low dew point double wheel dehumidifier structure, such as Figure 2 As shown in FIG, the key position of the low dew point dual-rotor dehumidifier (i.e., the sensor layout position) is taken as the node, the air circulation channel is taken as the edge, and the position relationship undirected graph is constructed according to the position relationship of the key position, as shown in FIG. Figure 3 As shown in the figure, the position relationship undirected graph G is defined as G = (V, E, A), where V is a set of N sensor nodes, E is the set of edges connecting these nodes, and A is the corresponding undirected graph adjacency matrix, where the element A[i, j] in the matrix satisfies A[i, j] = 1, indicating that the node V exists. i To Node V j If A[i,j]=0, it means that the node V i To Node V j There is no connection, so a corresponding position relationship undirected graph is generated based on the topological structure information of the low dew point dual-rotor dehumidifier.

[0045] S3: The historical operation status dataset of the dehumidifier is preprocessed according to the sampling time, node location, and data type to obtain a historical operation status database; the main purpose of data preprocessing is to convert the original collected data into data that can be used by the neural network model, that is, data standardization and vectorization, so as to achieve efficient data propagation in the model.

[0046] S3 is specifically:

[0047] After data cleaning, data standardization, node feature reconstruction and data vectorization of the historical operation status data, a historical operation status database is obtained.

[0048] Data cleaning includes filling missing data values ​​and correcting outliers. Through the physical discrimination method, missing data segments are eliminated, and mean interpolation is adopted for a small number of missing values ​​and outliers in some time steps. The data cleaning mainly includes eliminating missing data segments through the physical discrimination method, and mean interpolation is adopted for a small number of missing values ​​and outliers in some time steps. Note that the spatiotemporal graph attention neural network model captures the spatial correlation of the original data by node aggregation. For the situation where some types of sensors at some data collection points are not deployed or fail, they can be not processed in this step, but the missing node features can be updated by aggregating the features of neighboring nodes using the GAT network. Of course, this will affect the prediction accuracy of the model to a certain extent.

[0049] The data standardization process uses the Min-Max normalization method to map the original data X of the historical operation status data set to the [0,1] interval to obtain the normalized data X * , where the minimum value X of the training set, validation set, and test set min and the maximum value X max The training set X is usedmin Value and X max The value is as follows:

[0050]

[0051] The historical operation status database is composed of all data obtained after preprocessing, which should include the sensor collection data of each time slice after preprocessing and the recorded equipment control setting parameters, and the database should be filled and updated according to actual requirements to ensure the prediction accuracy.

[0052] Node feature reconstruction includes:

[0053] The air state data after data cleaning and data standardization are recorded as the first F-1 features at that moment. If the acquisition location is a cooler or a regenerative heater, that is, as a control node, the set temperature value of the cooler or the regenerative heater is taken as the Fth feature at that moment. Otherwise, the acquisition location is a non-control node, and the Fth feature at that moment is set to 0, thereby reconstructing the feature dimension of the acquisition location to F.

[0054] Data vectorization includes:

[0055] For the data after data cleaning, data standardization and node feature reconstruction, the reconstructed single-time operation status data is sequentially constructed into an operation status input matrix according to the time series relationship based on the time label, wherein the operation status input matrix is ​​specifically a data matrix formed by matching the data after node feature reconstruction according to the data and data point positions and data features. Then, according to the node position label, the position relationship undirected graph obtained in S2 is fused with the operation status input matrix to obtain a vectorized input matrix. Specifically, the fusion of the position relationship undirected graph G and the operation status input matrix is ​​to match the features in the operation status input matrix according to the position relationship, and obtain the vectorized input matrix at time t as follows: in Vectorize the features of the running status data. It represents the node V i The feature at time t has a feature dimension of F. Finally, the historical operation status database is composed of the vectorized input matrix at each time.

[0056] S4: Construct training set, validation set, and test set based on the historical operation status database;

[0057] S4 is specifically:

[0058] According to the size of the historical operation status data set, the time slice data in the historical operation status database that has been transferred to the vectorized input matrix is ​​divided into training set, validation set, and test set in a ratio of 8:1:1, which are used for model training, hyperparameter adjustment, and model performance testing respectively;

[0059] Construct training sets, validation sets, and test sets based on several vectorized input matrices, including:

[0060] The historical operation database is divided into training set, validation set and test set according to a certain ratio. The specific division method is: the first 80% of the historical data is used as the training set, the next 10% of the historical data is used as the validation set, and the last 10% of the historical data is used as the test set. In general, a part of the samples can be uniformly randomly sampled from the training set as the validation set, and the training set, validation set and test set all contain several vectorized input matrices. A sliding window with a set step size is used to slide each vectorized input matrix to obtain the input matrix of the spatiotemporal graph attention neural network. The input matrix of the spatiotemporal graph attention neural network is expressed as X in ={X t-T+1 ,…,X t-1 ,X t}, where T is the input sliding time window step size for spatiotemporal graph attention network training and prediction, X i is the vectorized input matrix at time i, that is, the historical status operation data in the past T time moments are used to predict the equipment operation status data in the future T′ time moments, where T′ is the length of the output sliding time window.

[0061] S5: Construct a spatiotemporal graph attention neural network model based on an undirected graph structure. After training the spatiotemporal graph attention neural network model based on an undirected graph structure using the training set, obtain a device operation status prediction model to capture the spatiotemporal dependency of data and the characteristics of data change trends.

[0062] In S5, the spatiotemporal graph attention neural network model based on the undirected graph structure includes a graph attention network GAT, a bidirectional gated recurrent unit BiGRU and a transformer network layer connected in sequence, and a multi-layer fully connected network for model output, such as Figure 4 As shown in the figure, the graph attention network GAT is used to read the spatial correlation of time series data on a time slice by time slice basis, the bidirectional gated recurrent unit BiGRU is used to fuse past information and future information to the prediction moment on a node by node basis, the time series Transformer layer is used to fuse the sliding time window input data information on a node by node basis to capture the global time correlation of the data, and the output layer is used to transform the output of the time series Transformer layer into the target output dimension to achieve the desired prediction.

[0063] Specifically:

[0064] The graph attention network GAT adopts aggregation operation and introduces the attention mechanism. By calculating the node feature similarity node by node, the features of neighboring nodes are aggregated to the central node to update the node features. It is applicable to multiple types of graph structures such as undirected graphs, directed graphs, and dynamic graphs. Among them, Global graph Attention can complete inductive learning tasks independently of the graph structure by performing attention calculation on any vertex in the entire graph, including situations where the graph structure is unknown or partially unknown; Mask graphAttention, which is the operation method adopted by the embodiment of this technology, limits the attention calculation to neighboring nodes, and performs transductive learning tasks based on prior graph structure knowledge, with higher prediction accuracy.

[0065] The graph attention network GAT stacks the Graph Attention Layer using masked self-attention to vectorize the input matrix X at a single moment. t , calculate the central node V i With neighboring node V j The correlation of the features, the neighbor nodes N i The features are aggregated to the central node to capture the spatial correlation of the running status data at each moment, which can be formally described as:

[0066]

[0067] in, is the central node V at the current time t i The feature representation, N i The central node V i The set of neighbor nodes of is the central node V at the current time t j The characteristic representation of V j ∈N i , W∈R F′×F is the linear transformation weight matrix of shared parameters, where F is The feature dimension of F′ is the feature output of the GAT layer. The feature dimension of V is [·∥·], which represents the matrix concatenation operation. i and node V j The transformed features are concatenated to obtain high-dimensional features. is the parameterized representation of the weight vector of the attention mechanism, mapping the obtained high-dimensional features to a real number, T Represents the matrix transpose operation, that is, the attention coefficient can be calculated And the attention coefficient e ij,t The normalized attention coefficient α is obtained through the softmax operation ij,t ,Right now LeakyReLU(·) is the activation function.

[0068] The current time t is the central node V i The characteristic output It can be calculated by the following formula:

[0069]

[0070] Among them, α ij,t is the normalized attention coefficient, Represents the neighboring node V j The features of , W is the weight matrix, and σ(·) is the nonlinear activation function.

[0071] Using the multi-head attention mechanism, the feature outputs obtained by K attention heads are concatenated to obtain the multi-head attention feature output:

[0072]

[0073] Among them, the feature output of the kth attention head is expressed as represents matrix concatenation, that is is the normalized attention coefficient calculated for the kth attention head, W k is the kth attention head weight matrix, and Concat() means concatenation.

[0074] The final attention layer of the graph attention network GAT is concatenated with an average substitution matrix and finally applies a nonlinear activation function σ(·), which is formally described as:

[0075]

[0076] The output matrix of the GAT network is X′=(X′ t-T+1 ,…,X′ t-1 ,X′ t ), X′∈R F′×N×T , where the potential feature output matrix at each moment is

[0077] In practice, the GAT graph attention network can add self-loop edges according to the attributes of the dataset, task type, and target output dimension. That is, it adds loop edges connecting the central node to itself on the original graph structure, and adds self-embedding operations to the node feature aggregation operations, so that the network can decide the learning weights of the node information itself. This can effectively reduce the noise impact brought by node embedding and increase prediction accuracy in practice.

[0078] The BiGRU bidirectional gated recurrent unit is a variant of the GRU gated recurrent unit. It is usually composed of two unidirectional GRUs structures with the same input and opposite directions. It can capture the forward propagation information and backward propagation information of the time series data respectively. The actual output implicit state can be obtained by aggregating the implicit output states of the two GRUs structures.

[0079] The BiGRU bidirectional gated recurrent unit is composed of two unidirectional GRUs in opposite directions, and the number of GRU units in each unidirectional GRU is determined by the sliding window size T, that is, the entire network has 2T GRU units, which are used to process data forward and backward in time series, and capture the forward and backward information of sequence data at the same time, which can be used to capture the impact of past and future moments on current moment data;

[0080] The simplified GAT network is simplified to model f GAT (·), the GAT network output matrix at the current time t is X′ t =f GAT (A,X t ), by X′ t The BiGRU network input matrix is ​​obtained by matrix transposition Or it can be expressed as That is, node V i The feature vector at time t represents The input matrix obtained by transposition and vector concatenation is known to have the hidden representation H at the previous time t-1 t-1 , then the calculation of a single GRU can be formally expressed as:

[0081]

[0082] The operation of a single GRU unit is independent of each node, and the GRU parameters of all nodes are shared, where H t [i,:] is the node V at the current time t i The implicit output state of is the candidate hidden state at time t, H t-1 [i,:] is the node V at the current time t-1 i The implicit output state of is the candidate hidden state at time t-1, z t and r t are the update gate and reset gate of the GRU at the current time t, W z , W r , W h For GRU with respect to the input matrix The three learnable weight matrices, U z , U r and U hGRU is the candidate hidden state of the previous layer The three learnable weight matrices, b z , b r , b h are the three learnable bias parameters of GRU, σ r (·) is a nonlinear activation function, ⊙ is an element-wise product operation;

[0083] The output of BiGRU consists of two unidirectional GRUs, and the transposed GAT network output matrix As the input of two GRUs at the same time, the output of the forward propagation GRU is Indicates that the output of the back-propagation GRU is given by Indicates that the output H of BiGRU BiGRU ∈R N×F′×T , which can be formally expressed as:

[0084]

[0085] H BiGRU =(H t-T+1 ,…,H t-1 ,H t )

[0086] Among them, H t is the output matrix of the BiGRU network at time t, W f and W b are the learnable weight matrices of the forward and back-propagation GRUs, respectively, and B is the learnable weight of the set item.

[0087] The Time Series Transformer has the ability to model long-term and short-term time series features, encode sequence position information, convert multi-time slice data input matrix into multi-node data matrix, integrate Scaled Dot-ProductAttention and Multi-Head Attention mechanisms, perform feature fusion on BiGRU's multi-time slice implicit output state matrix, and extract long-term temporal correlation of time series data.

[0088] The Transformer layer consists of a multi-head attention layer, a residual connection and layer normalization layer, and a shared feedforward neural network layer. BiGRU Position encoding, feature concatenation, and multi-head attention fusion are performed to capture the global temporal dependency of the data.

[0089] The multi-head attention mechanism of the Transformer layer treats different positions equally when calculating the attention function. To ensure that the Transformer layer can effectively capture H t[i,:] Relative position in the sequence, using position encoding e t , the formula is as follows:

[0090]

[0091] Among them, d model The dimension of the position encoding embedding space can be determined according to the sliding window size. is the node V after position encoding i The output implicit state matrix at time t, H e ∈R N×F′×T is the actual input matrix of the Transformer layer;

[0092] The multi-head attention mechanism of the Transformer layer is achieved by feature splicing, that is, splicing the same node V in chronological order. i The row vector representation of each moment Get the node matrix And the node matrix Linear mapping to the index matrix Query, address matrix Key, and value matrix Value is formally expressed as:

[0093]

[0094] in Represents node V i The index matrix of Represents node V i The address matrix, Represents node V i The value matrix of For the three linear mapping weight matrices shared by the parameters, the attention function It can be formalized as:

[0095]

[0096] Among them, softmax() represents a classifier.

[0097] The multi-head attention mechanism consists of K attention heads, using K sets of projection matrices to Projected to K different groups of indexes, keys, and values, where the output of the kth attention head is denoted as Head k , concatenate the output matrices of K attention heads and aggregate the output representations of different attention heads Enhance the expressiveness of the model, formalized as follows:

[0098]

[0099] in, is the linear mapping matrix of the kth attention head, W O Output matrix for learnable linear mapping;

[0100] The output state of the multi-head attention layer is transmitted point by point to the feedforward neural network layer, and after residual connection and layer normalization, the output matrix of the Transformer layer is obtained.

[0101] Furthermore, the output of the spatiotemporal graph attention neural network model is given by a prediction layer composed of a multi-layer feedforward neural network, represented by H O As the input of the prediction layer, it is expressed as:

[0102]

[0103] The output layer is constructed by O The output matrix is ​​obtained by passing through a multi-layer fully connected feedforward network node by node. where F O is the node feature dimension output by the neural network, T′ is the length of the sliding window used by the neural network for prediction, that is, the output matrix is ​​the F of all nodes V in the future T′ moments O Heavy features;

[0104] Furthermore, the training of the spatiotemporal graph attention neural network is to split the historical running status data X into X train , X valid , X test As training set data, validation set data and test set data respectively, in the training data X train A window of length T+T′ is slid upward to generate training sequences, where the first T sequences are used as input and the remaining T′ sequences are used as basic facts, with the mean absolute error MAE as the loss, formally expressed as:

[0105]

[0106] in, is the output prediction value of the model at time ti, X ti is the true value of the model at time ti, ti = t+T+1, t+T+T′, t+1, t+T, t+T+i, f(·) is the prediction model trained by minimizing the loss function, that is, the equipment operation status prediction model, Loss is the loss value, || represents the absolute value, and d() represents the average absolute error between the predicted value and the true value output at each time.

[0107] S6: Generate a vectorized input matrix based on the real-time collected air state data and equipment control parameters, and input the vectorized input matrix into the equipment operation state prediction model. The model outputs the predicted equipment operation state at the future moment and the optimized equipment control parameters of the surface cooling and regeneration heater, that is, the characteristics of each node in the graph structure at each future moment. The pre-processed T neural network input matrices of the past and current time slices are input into the pre-trained deep learning fusion model to obtain the air state prediction value of each node at the future moment. The corresponding control node will also obtain the corresponding controller set temperature prediction value. The air state change trend in the equipment can be judged based on the air state prediction value of each node. For PID controlled equipment, the set parameter adjustment can be pre-set based on the prediction value. The specific set parameter value can be assisted by the air state prediction value and the set temperature prediction value. In the specific practice process, it can effectively improve the equipment response speed, reduce the equipment energy consumption, and achieve the purpose of energy saving optimization.

[0108] On the one hand, the present invention also proposes a low dew point dual-rotor dehumidifier energy-saving prediction system based on spatiotemporal graph attention, comprising:

[0109] The operation status data acquisition module is used to collect the air status data and equipment control parameters at key locations where the air status of the low dew point dual-rotor dehumidifier is significantly different, such as the air inlet, air outlet, and mixed air, in real time, so as to construct an operation status data set;

[0110] A position relationship undirected graph generation module is used to generate a position relationship undirected graph according to the topological relationship of the low dew point dual-rotor dehumidifier structure;

[0111] A data preprocessing module, used to convert the operating status data set into a vectorized input matrix;

[0112] The equipment operation status prediction module is used to store the equipment operation status prediction model and predict the status of the low dew point dual-rotor dehumidifier according to the input vectorized input matrix to obtain the predicted equipment operation status at the future moment and the optimized equipment control parameters.

[0113] Finally, it should be noted that the above embodiments and explanations are only used to illustrate the technical solution of the present invention rather than to limit it. Those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope disclosed in the technical solution of the present invention, which should be included in the scope of protection of the claims of the present invention.

Claims

1. A low dew point double-rotor dehumidifier energy-saving prediction method based on spatiotemporal graph attention, characterized in that: The following steps are involved: S1: Sensors are placed at key positions of the low dew point dual-rotor dehumidifier to collect air status data and equipment control parameters of the surface cooling and regeneration heater at the above key positions in real time, and the dehumidifier historical operation status data set is constructed based on the air status data and equipment control parameters of the surface cooling and regeneration heater collected in real time by the sensors; S2: Taking the layout positions of the low dew point dual-rotor dehumidifier sensors as nodes and the air circulation channels inside the rotary dehumidifier as edges, an undirected position relationship graph is constructed based on the relative relationship between the sensor layout positions and the air circulation channels. S3: preprocess the historical operation status data set of the dehumidifier according to sampling time, node location, and data type to obtain a historical operation status database; S4: Construct a neural network training set based on the historical operation status database; S5: Construct a spatiotemporal graph attention neural network model based on an undirected graph structure, and use the neural network training set to train the spatiotemporal graph attention neural network model based on an undirected graph structure to obtain a trained equipment operation status prediction model; S6: Generate a vectorized input matrix based on the real-time collected air state data and equipment control parameters, process the vectorized input matrix using a sliding window, and then input it into the trained equipment operation state prediction model. The output of the model includes the equipment operation state at the future moment and the optimized equipment control parameters of the surface cooling and regeneration heater.

2. The energy-saving prediction method for low dew point dual-rotor dehumidifier based on spatiotemporal graph attention according to claim 1 is characterized in that: In S1, the air state data includes air temperature, humidity, air volume or other collectible air state data, and the equipment control parameters are set temperature values ​​of the surface cooler and the regeneration heater.

3. The energy-saving prediction method for low dew point dual-rotor dehumidifier based on spatiotemporal graph attention according to claim 1 is characterized in that: The S3 is specifically: The historical operation status data is cleaned, standardized, reconstructed and vectorized at each sampling moment to obtain a historical operation status database.

4. The energy-saving prediction method for low dew point dual-rotor dehumidifier based on spatiotemporal graph attention according to claim 3 is characterized in that: In S3, node feature reconstruction includes: For the data after data standardization, the air state data at each moment is recorded as the first F-1 features of that moment. If the collection position is the air outlet position through the cooler or regeneration heater, the set temperature value of the cooler or regeneration heater is used as the Fth feature of that moment. Otherwise, the Fth feature of that moment is set to 0, thereby reconstructing the feature dimension of the collection position to Fth.

5. The energy-saving prediction method for low dew point dual-rotor dehumidifier based on spatiotemporal graph attention network according to claim 3 is characterized in that: In S3, data vectorization includes: For the data reconstructed by node features, the reconstructed single-moment operating status data is constructed into an operating status input matrix in sequence according to the time label; then, according to the node position label, the position relationship undirected graph and the operating status input matrix are fused to obtain a vectorized input matrix.

6. The energy-saving prediction method for low dew point dual-rotor dehumidifier based on spatiotemporal graph attention according to claim 1 is characterized in that: The process of processing the vectorized input matrix by using a sliding window includes: The vectorized input matrices are respectively processed by sliding using a sliding window, thereby obtaining several input matrices of the same size.

7. The energy-saving prediction method for low dew point dual-rotor dehumidifier based on spatiotemporal graph attention according to claim 1 is characterized in that: In S5, the spatiotemporal graph attention neural network model based on an undirected graph structure includes a graph attention network, a bidirectional gated recurrent unit and a temporal Transformer network layer connected in sequence, and a multi-layer fully connected network for model output.

8. A low dew point double-rotor dehumidifier energy-saving prediction system based on spatiotemporal graph attention, characterized in that: include: The operation status data collection module is used to collect the air status data and equipment control parameters at the key positions of the low dew point double-rotor dehumidifier in real time, so as to construct the operation status data set; A position relationship undirected graph generation module is used to generate a position relationship undirected graph according to the topological relationship of the low dew point dual-rotor dehumidifier structure; A data preprocessing module, used to convert the operating status data set into a vectorized input matrix; The equipment operation status prediction module is used to store the equipment operation status prediction model and predict the status of the low dew point dual-rotor dehumidifier according to the input vectorized input matrix, so as to obtain the predicted equipment operation status at the future moment and the optimized equipment control parameters of the surface cooler and the regeneration heater.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of a low dew point dual-rotor dehumidifier energy-saving prediction method based on spatiotemporal graph attention as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a low dew point dual-rotor dehumidifier energy-saving prediction method based on spatiotemporal graph attention as described in any one of claims 1 to 7 are implemented.

Images