Space-time joint temperature prediction method based on deep reversible network

Abstract

The invention belongs to the technical field of temperature prediction, and particularly discloses a space-time joint temperature prediction method based on a deep reversible network. A prediction model used in the method comprises a pixel shuffling layer, a forward encoder, a reverse decoder and a long and short term integrated prediction module used for bridging the forward encoder and the reverse decoder. An input image is subjected to shape transformation through one pixel shuffling layer, the forward encoder is used for extracting deep features, the long-short term integrated prediction module is used for further extracting spatial-temporal features from the deep features, the reverse decoder is used for decoding the spatial-temporal features, and the resolution of the decoded features is restored to be the same as that of the input image through the other pixel shuffling layer; the forward encoder and the reverse decoder have the same structure and share parameters, and each of the forward encoder and the reverse decoder comprises a plurality of multi-level reversible affine coupling structures; the long and short term integrated prediction module comprises a plurality of long and short term integrated prediction units, and the output feature of the current long and short term integrated prediction unit and the input feature of the previous long and short term integrated prediction unit are added to serve as the input feature of the next long and short term integrated prediction unit. According to the method, feature extraction can be carried out more flexibly, information loss is minimized, spatio-temporal information is fully reserved, and prediction precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of temperature prediction, and specifically is a spatiotemporal joint temperature prediction method based on a deep reversible network. Background Art

[0002] Temperature spatiotemporal prediction refers to predicting the temperature changes in a specific area over a period of time in the future. Temperature data has a strong spatiotemporal correlation, so fully modeling the correlation between time and space dimensions is the key to achieving accurate temperature prediction.

[0003] At present, the temperature spatiotemporal prediction methods can be roughly divided into two categories. The first category is a physical-based numerical simulation method. According to relevant physical and chemical theories, a complex mathematical model is established by collecting relevant data, and then a high-performance computer is used to simulate the dynamic changes of temperature in space to achieve the purpose of prediction. The second category is a data-driven deep learning method, which has two main ideas. One is a complex model based on a single-input single-output (SISO) structure, such as a recurrent neural network, a convolutional long short-term memory, and its variants. The model learns the hidden state of historical information and generates the next frame based on the current frame prediction. When the generated frame becomes longer, the error continues to accumulate. Small errors can easily be magnified into serious compound errors over time, resulting in a rapid decrease in the quality and accuracy of the subsequent generated frames. The other is a model based on a multiple-input multiple-output (MIMO) structure, which inputs all known frames into the encoder and outputs all unknown frames from the decoder at one time.

[0004] Existing deep learning-based spatiotemporal temperature prediction methods focus more on how to better quantify the difference between real frames and predicted frames. However, it is also important to make better use of the intra-frame and inter-frame relationships of image frames for prediction, so that the model can make better use of long-term and short-term spatiotemporal information, and further retain spatiotemporal information during feature extraction, reduce memory requirements and computational costs, and ensure that no information is lost to the greatest extent.

[0005] In deep learning or probabilistic modeling, reversibility is generally a useful property that allows data to be transformed without losing information, thereby solving the problem that traditional neural networks compress and map input data during forward propagation, resulting in partial information loss. Reversible networks allow input data to be restored through reverse operations, ensuring that complete information is retained at every step of the entire network, which is very important for tasks with high precision requirements such as temperature prediction. Avoiding information loss helps to improve the accuracy of prediction results. Based on the above problems, the present invention proposes a spatiotemporal joint temperature prediction method based on a deep reversible network. Summary of the invention

[0006] In view of the deficiencies in the prior art, the technical problem that the present invention intends to solve is to provide a spatiotemporal joint temperature prediction method based on a deep reversible network.

[0007] The present invention solves the technical problem by adopting the following technical solution: A spatiotemporal temperature prediction method based on a deep reversible network, characterized in that it comprises the following steps: Step 1: Obtain the temperature image of a certain area over a period of time, standardize the temperature image, and obtain a data set; Step 2: Build a prediction model based on a deep reversible network, including a pixel shuffle layer, a forward encoder, a reverse decoder, and a long-term and short-term integrated prediction module used to bridge the forward encoder and the reverse decoder; the input image undergoes shape transformation through a pixel shuffle layer, the forward encoder is used to extract deep features, the long-term and short-term integrated prediction module further extracts spatiotemporal features from the deep features, the reverse decoder is used to decode the spatiotemporal features, and the decoded features undergo another pixel shuffle layer to restore the resolution to the same as the input image; The forward encoder and the reverse decoder have the same structure and share parameters, both of which include multiple multi-level reversible affine coupling structures; The process of extracting features from a multi-level reversible affine coupling structure is expressed as: (1) (2) In the formula, and Indicates the forward encoder Input features of a multi-level reversible affine coupling structure; , Indicates the forward encoder The output characteristics of the multi-level reversible affine coupling structure, namely Input features of a multi-level reversible affine coupling structure; represents the Hadamard product, represents the exponential function, , represents the convolution operation that generates the scaling factor, , represents the convolution operation that generates the translation bias; Reverse Decoder The process of decoding the multi-level reversible affine coupling structure features is expressed as: (14) (15) In the formula, and Reverse decoder The output characteristics of the multi-level reversible affine coupling structure, namely The input features of a multi-level reversible affine coupling structure, , Reverse decoder Output features of a multi-level reversible affine coupling structure; Step 3: Train and optimize the prediction model, and use the trained prediction model for temperature prediction.

[0008] Furthermore, the long-term and short-term integrated prediction module includes a plurality of long-term and short-term integrated prediction units, and the output feature of the current long-term and short-term integrated prediction unit is added to the input feature of the previous long-term and short-term integrated prediction unit as the input feature of the next long-term and short-term integrated prediction unit; The long-term and short-term integrated prediction unit includes a short-term dependency block, a long-term and short-term aggregation block and an information update block; the prediction process of the short-term dependency block of the l-th long-term and short-term integrated prediction unit is expressed as: (3) (4) (5) (6) In the formula, , represents the trend gate and update gate of the short-term dependency block of the l-th long-term integrated prediction unit at time step t, It is the feature obtained by adding the spatiotemporal features predicted by the l-1th long-term and short-term integrated prediction unit at time step t and the input features of the l-2th long-term and short-term integrated prediction unit at time step t. is the spatiotemporal feature of the memory unit extracted by the l-1th long-term and short-term integrated prediction unit at time step t, , represents the weight matrix of the trend gate, , represents the weight matrix of the update gate, represents the hidden state features of the short-term dependency block of the l-th long-term integrated prediction unit at time step t, , The weight matrix representing the hidden state; represents the output features of the short-term dependency block of the l-th long-term and short-term integrated prediction unit at time step t, that is, the short-term spatiotemporal dependency features extracted by the l-th long-term and short-term integrated prediction unit at time step t; represents the convolution operation, Indicates activation operation. Indicates the number of frames contained in the time step; The prediction process of the long-term and short-term aggregation block of the l-th long-term and short-term integrated prediction unit is expressed as: (7) (8) In the formula, represents the long-term spatiotemporal dependency features obtained by the l-th long-term and short-term integrated prediction unit at time step t, represents the spatiotemporal characteristics of the memory units from the l-1-kth to the l-1th long-term and short-term integrated prediction units at time step t, represents the attention mechanism, represents the mixed spatiotemporal dependency features obtained by the l-th long-term and short-term integrated prediction unit at time step t; The update process of the information update block of the l-th long-term and short-term integrated prediction unit is expressed as: (9) (10) (11) (12) (13) In the formula, , represents the weight matrix and bias of the input gate, represents the input gate of the information update block of the l-th long-term and short-term integrated prediction unit at time step t, represents the fusion features of the information update block of the l-th long-term and short-term integrated prediction unit at time step t, , Represents the weight matrix and bias of feature fusion, , represents the weight matrix and bias of the output gate, represents the output gate of the information update block of the l-th long-term and short-term integrated prediction unit at time step t, It is the spatiotemporal feature predicted by the lth long-term and short-term integrated prediction unit at time step t.

[0009] Furthermore, the input image is transformed from 1×h×w to 4×h / 2×w / 2 through a pixel shuffling layer, and the transformed image is split into two images of shape 2×h / 2×w / 2 as the input of the forward encoder; where h and w represent the height and width of the image, respectively.

[0010] Furthermore, during the prediction model training and optimization process, the training loss is calculated according to the following objective optimization function; (16) In the formula, is the root mean square error, is the number of samples, , It is The true value and predicted value of each sample.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention constructs a prediction model based on a reversible network. The forward encoder and the reverse decoder form a parameter-sharing non-volume-preserving reversible bidirectional autoencoder, which can extract features more flexibly, minimize information loss, and fully retain spatiotemporal information. The reversible bidirectional autoencoder not only greatly reduces memory requirements and computing costs, but also theoretically guarantees the characteristics of no information loss, and extends reversibility from the spatial domain to the temporal domain.

[0012] 2. LSTM-based memory networks have a strong ability to capture long-term and short-term dependency information and have been widely used in most spatiotemporal prediction models. However, the traditional convolution of LSTM relies heavily on convolution kernels to model relevant structural information. Its convolution kernel is a building block that processes a local neighborhood within a fixed-size receptive field (usually 3×3 or 5×5). This local fixed-size convolution kernel is inefficient when processing structural information in large areas, and because the current end-to-end model is limited by the length of the input sequence, it often encounters the problem of spatiotemporal illusion, that is, similar input time series are followed by different future values, and vice versa. Therefore, the present invention proposes a long-term and short-term integrated prediction unit to predict short-term spatiotemporal dependencies in short-term dependency blocks. Through the learned short-term change information and the long-term change information aggregated from the memory unit information of multiple time steps in the past, the gating mechanism is used to adaptively update the memory unit and output features so that the spatiotemporal attention can capture the dependencies in the spatial and temporal domains. By capturing the relative changes in local and global spatiotemporal space, the prediction features can be generated more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the structure of the prediction model of the present invention; Figure 2 It is a principle diagram of the multi-level reversible affine coupling structure of the present invention; Figure 3 It is a schematic diagram of the long-term and short-term integrated prediction unit of the present invention. DETAILED DESCRIPTION

[0014] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods, but the protection scope of the present application is not limited thereto.

[0015] The present invention provides a spatiotemporal temperature prediction method based on a deep reversible network (hereinafter referred to as the method, see Figures 1 to 3 ), including the following steps: Step 1: Obtain temperature images, process the temperature images, and form a data set; The historical temperature images of a certain area over a period of time are obtained, the historical temperature images are standardized, and a sliding window is used to slice them in chronological order to obtain several temperature images, which constitute the data set.

[0016] Step 2: Build a prediction model based on a deep reversible network. The prediction model is based on the probability chain rule and models the joint distribution by decomposing it into conditional products according to a fixed dimensional order, which simplifies log-likelihood evaluation and sampling.

[0017] like Figure 1 As shown in Figure 1, the prediction model includes a pixel shuffle layer, a forward encoder, a reverse decoder, and a long-term and short-term integrated prediction module for bridging the forward encoder and the reverse decoder. In order to achieve the reversibility of the prediction model, the input temperature image After a pixel shuffling layer, the shape is transformed from 1×h×w to 4×h / 2×w / 2, and the transformed image is split into two images with a shape of 2×h / 2×w / 2. and ; Where 1 represents a single channel, h and w represent the height and width of the image respectively; the forward encoder is used to and The deep features are extracted from the long-term and short-term integrated prediction module, and the spatiotemporal information is retained during the feature extraction process; the long-term and short-term integrated prediction module is used to further obtain the dependencies in the spatial and temporal dimensions to extract spatiotemporal features, and the reverse decoder decodes the spatiotemporal features extracted by the long-term and short-term integrated prediction module to obtain the predicted image. and ; Prediction image and After another pixel shuffling layer, the channel information is restored to the spatial dimension, so that the resolution of the predicted image is restored to the same as the input image, and the predicted temperature image is obtained. .

[0018] The forward encoder and the reverse decoder have the same structure, both of which are stacked with multiple multi-level reversible affine coupling structures, and the parameters of the forward encoder and the reverse decoder are shared. and The input is sent to the forward encoder, and features are extracted layer by layer through multiple multi-level reversible affine coupling structures. The output of the last multi-level reversible affine coupling structure is the output of the forward encoder. The process of extracting features from a multi-level reversible affine coupling structure is expressed as: (1) (2) In the formula, and Indicates the forward encoder Input features of a multi-level reversible affine coupling structure; , Indicates the forward encoder The output characteristics of the multi-level reversible affine coupling structure, namely Input features of a multi-level reversible affine coupling structure; represents the Hadamard product, represents the exponential function, , represents the convolution operation that generates the scaling factor, , Represents a convolution operation that generates a translation bias.

[0019] The convolution operation that generates the scaling factor multiplies a learnable scaling factor on each dimension of the feature to control the amplitude of the output feature; the convolution operation that generates the translation bias adds a learnable translation bias on each dimension of the feature to adjust the center position of the output feature; by introducing these learnable parameters, the model can dynamically adjust the intensity of the nonlinear transformation, thereby improving the convergence rate and avoiding problems such as gradient disappearance as much as possible.

[0020] For multi-level reversible affine coupling structures of different levels, the size of input features can be selectively reduced through convolution operations, for example, the size is processed from 4×h / 2×w / 2 to 16×h / 4×w / 4, and the data is processed at different granularities in order to better capture the hierarchical structure and detail information of the data, thereby improving the model's ability to model data distribution. Multi-level reversible affine coupling structures of different levels can learn the feature representation of data at different levels. The low-level multi-level reversible affine coupling structure captures more pixel-level detail information, while the high-level multi-level reversible affine coupling structure can capture more abstract semantic information, thereby capturing global structure and local details at the same time, thereby generating higher quality predicted images.

[0021] The input feature of the long-term and short-term integrated prediction module is the output feature of the positive phase encoder, which is recorded as and The long-term and short-term integrated prediction module includes multiple long-term and short-term integrated prediction units. The output features of the current long-term and short-term integrated prediction unit are added to the input features of the previous long-term and short-term integrated prediction unit as the input features of the next long-term and short-term integrated prediction unit. The input features of the first long-term and short-term integrated prediction unit are , the output features of the first long-term and short-term integrated prediction unit are After addition, they are used as input features of the second long-term and short-term integrated prediction unit. The long-term and short-term integrated prediction module adopts an architecture similar to the forward encoder and reverse decoder to ensure pixel alignment between input features and output features, that is, each feature position can be traced back to a certain pixel, making it compatible with the forward encoder and reverse decoder.

[0022] The long-term and short-term integrated prediction unit includes a short-term dependency block, a long-term and short-term aggregation block, and an information update block; compared with the LSTM-based prediction module, each long-term and short-term integrated prediction unit has only five gates, namely, a trend gate, an update gate, an input gate, a forget gate, and an output gate. The short-term dependency block uses the trend gate to model the short-term spatiotemporal trend information, and uses the update gate to update the current spatiotemporal information to the future short-term spatiotemporal information, so that the short-term dependency block can predict the short-term spatiotemporal dependency. The long-term and short-term aggregation block is used to aggregate the long-term change information from the past memory units, and the learned short-term change information to obtain the long-term and short-term mixed spatiotemporal dependency. The information update block uses a gating mechanism to adaptively update the memory unit spatiotemporal information and the mixed spatiotemporal dependency information, so that the long-term and short-term integrated prediction module can capture the dependencies in the spatial and temporal domains.

[0023] For the prediction of the l-th long-term and short-term integrated prediction unit at time step t, the input features include two, namely the spatiotemporal features predicted by the l-1-th long-term and short-term integrated prediction unit at time step t The feature obtained by adding the input feature of the l-2th long-term and short-term integrated prediction unit at time step t , the spatiotemporal characteristics of the memory unit obtained by the l-1th long-term and short-term integrated prediction unit at time step t The prediction process of the short-term dependency block of the l-th long-term integrated prediction unit is expressed as: (3) (4) (5) (6) In the formula, , represents the trend gate and update gate of the short-term dependency block of the l-th long-term integrated prediction unit at time step t, , represents the weight matrix of the trend gate, , represents the weight matrix of the update gate, represents the hidden state features of the short-term dependency block of the l-th long-term integrated prediction unit at time step t, , The weight matrix representing the hidden state; Represents the output features of the short-term dependency block of the l-th long-term and short-term integrated prediction unit at time step t, that is, the short-term spatiotemporal dependency features obtained by the l-th long-term and short-term integrated prediction unit at time step t; represents the convolution operation, Indicates activation operation. Indicates the number of frames contained in the time step; The prediction process of long-term and short-term aggregation blocks is expressed as: (7) (8) In the formula, represents the long-term spatiotemporal dependency features obtained by the l-th long-term and short-term integrated prediction unit at time step t, represents the spatiotemporal characteristics of the memory units from the l-1-kth to the l-1th long-term and short-term integrated prediction units at time step t, represents the attention mechanism, represents the mixed spatiotemporal dependency features obtained by the l-th long-term and short-term integrated prediction unit at time step t; The update process of the information update block is expressed as: (9) (10) (11) (12) (13) In the formula, , represents the weight matrix and bias of the input gate, represents the input gate of the information update block of the l-th long-term and short-term integrated prediction unit at time step t, represents the fusion features of the information update block of the l-th long-term and short-term integrated prediction unit at time step t, , Represents the weight matrix and bias of feature fusion, , represents the weight matrix and bias of the output gate, represents the output gate of the information update block of the l-th long-term and short-term integrated prediction unit at time step t, It is the output feature of the lth long-term and short-term integrated prediction unit, that is, the spatiotemporal feature predicted at time step t.

[0024] The memory unit not only updates the local spatiotemporal dependency information through convolution operations, but also obtains the global spatiotemporal dependency information through attention. Contains more complete and comprehensive space-time information.

[0025] The reverse decoder implements the reverse process by inverse transformation and decodes the spatiotemporal features obtained by the long-term and short-term integrated prediction module. The process of decoding the multi-level reversible affine coupling structure features is expressed as: (14) (15) In the formula, and Reverse decoder The output characteristics of the multi-level reversible affine coupling structure, namely The input features of a multi-level reversible affine coupling structure, , Reverse decoder Output features of a multi-level reversible affine coupling structure.

[0026] Step 3: Use the data set to train and optimize the prediction model, and use the trained prediction model for spatiotemporal temperature prediction; In the process of training and optimization, the root mean square error is used as the target optimization function, and the training loss is measured by the target optimization function; the root mean square error can better reflect the actual situation of the prediction value error, and is very sensitive to large errors and small errors, and can well reflect the prediction accuracy. The smaller the root mean square error, the higher the prediction accuracy; (16) In the formula, is the root mean square error, is the number of samples, , It is The true value and predicted value of each sample.

[0027] In order to verify the effectiveness of the prediction model of the present invention, a test was conducted on the CSFR dataset, and a subset of the CSFR dataset was selected, which contained historical temperature images from January 1979 to December 2015, covering the area of ​​8°N-54°S and 80°W-25°W. In order to verify the prediction ability of the model for shorter and longer sequences, two prediction strategies were adopted. The first strategy uses a sliding window with a width of 10 frames to slice continuous temperature images. Each sequence consists of ten continuous frames, of which the first 5 frames are input frames and the last five frames are prediction frames, to verify the prediction ability of short sequences. The second strategy uses a 20-frame sliding window to slice continuous images, so each sequence consists of 20 continuous frames, of which the first 5 frames are input frames and the last 15 frames are prediction frames, to verify the prediction ability of long sequences. The number of sequence samples for the two strategies are 54047 and 54037 respectively, and the sample data shapes are 10×1×32×32 (the first strategy) and 20×1×32×32 (the second strategy). The value 1 in the data shape represents a single channel (similar to a grayscale image).

[0028] The prediction model of the present invention is compared with the existing eight prediction models (ConvLSTM [1] 、PredRNN [2] 、E3dLSTM [3] 、MIM [4] 、MAU [5] 、SimVP [6] 、MMVP [7] and TAU [8] ), and selected four indicators, MSE (root mean square error), MAE (mean absolute error), FLOPs (floating point operations, measuring algorithm complexity) and parameter quantity, to evaluate the model performance. The test results are shown in Table 1.

[0029]

[0030] The above test results show that the prediction model of the present invention has the best prediction effect for both short and long temperature image sequences, and has different degrees of improvement compared with the existing prediction models. This is because the use of parameter-sharing non-volume-preserving reversible bidirectional autoencoders makes the feature extraction process more flexible, minimizes information loss, and improves the prediction accuracy of temperature spatiotemporal series data, significantly reduces the prediction error, and has advantages in reducing the number of parameters and computational complexity. At the same time, the long-short-term integrated prediction unit can effectively handle short-term and long-term spatiotemporal dependencies, significantly improving the overall prediction performance of the model. The unit first aggregates past spatiotemporal information through short-term dependency blocks, and uses memory units to predict short-term dependencies, so that local changes in spatiotemporal sequences can be quickly captured and processed. Subsequently, the long-short-term aggregation block uses an information joint mechanism based on resolution-preserving blocks to fuse short-term dependency and long-term dependency features to model and predict complex mixed dependency spatiotemporal information. Finally, the information update block uses a gating mechanism to adaptively update the memory unit and output features. This strategy ensures that the model can not only cope with short-term spatiotemporal dynamic changes, but also effectively capture long-term dependent characteristics, thereby improving the prediction accuracy of temperature spatiotemporal series data and significantly reducing prediction errors.

[0031] The sources of the existing prediction models are: [1] Shi X, Chen Z, Wang H, et al. Convolutional LSTM network: Amachine learning approach for precipitation nowcasting[J]. Advances in neuralinformation processing systems, 2015, 28. [2] Wang Y, Long M, Wang J, et al. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms[J]. Advances in neuralinformation processing systems, 2017, 30. [3] Wang Y, Jiang L, Yang M H, et al. Eidetic 3D LSTM: A model forvideo prediction and beyond[C] / / International conference on learningrepresentations. 2018. [4] Wang Y, Zhang J, Zhu H, et al. Memory in memory: A predictiveneural network for learning higher-order non-stationarity from spatiotemporaldynamics[C] / / Proceedings of the IEEE / CVF conference on computer vision andpattern recognition. 2019: 9154-9162. [5] Chang Z, Zhang X, Wang S, et al. Mau: A motion-aware unit forvideo prediction and beyond[J]. Advances in Neural Information ProcessingSystems, 2021, 34: 26950-26962. [6] Gao Z, Tan C, Wu L, et al. Simvp: Simpler yet better videoprediction[C] / / Proceedings of the IEEE / CVF conference on computer vision andpattern recognition. 2022: 3170-3180. [7] Zhong Y, Liang L, Zharkov I, et al. Mmvp: Motion-matrix-basedvideo prediction[C] / / P-roceedings of the IEEE / CVF International Conference onComputer Vision. 2023: 4273-4283. [8] Tan C, Gao Z, Wu L, et al. Temporal attention unit: Towardsefficient spatiotemporal predictive learning[C] / / Proceedings of the IEEE / CVFConference on Computer Vision and Pattern Recognition. 2023: 18770-18782. Any matters not described in the present invention are applicable to the prior art.

Claims

1. A spatiotemporal temperature prediction method based on deep reversible network, characterized in that: The steps include: Step 1: Obtain the temperature image of a certain area over a period of time, standardize the temperature image, and obtain a data set; Step 2: Build a prediction model based on a deep reversible network, including a pixel shuffle layer, a forward encoder, a reverse decoder, and a long-term and short-term integrated prediction module used to bridge the forward encoder and the reverse decoder; the input image undergoes shape transformation through a pixel shuffle layer, the forward encoder is used to extract deep features, the long-term and short-term integrated prediction module further extracts spatiotemporal features from the deep features, the reverse decoder is used to decode the spatiotemporal features, and the decoded features undergo another pixel shuffle layer to restore the resolution to the same as the input image; The forward encoder and the reverse decoder have the same structure and share parameters, both of which include multiple multi-level reversible affine coupling structures; The process of extracting features from a multi-level reversible affine coupling structure is expressed as: (1) (2) In the formula, and Indicates the forward encoder Input features of a multi-level reversible affine coupling structure; , Indicates the forward encoder The output characteristics of the multi-level reversible affine coupling structure, namely Input features of a multi-level reversible affine coupling structure; represents the Hadamard product, represents the exponential function, , represents the convolution operation that generates the scaling factor, , represents the convolution operation that generates the translation bias; Reverse Decoder The process of decoding the multi-level reversible affine coupling structure features is expressed as: (14) (15) In the formula, and Reverse decoder The output characteristics of the multi-level reversible affine coupling structure, namely The input features of a multi-level reversible affine coupling structure, , Reverse decoder Output features of a multi-level reversible affine coupling structure; Step 3: Train and optimize the prediction model, and use the trained prediction model for temperature prediction.

2. The spatiotemporal temperature prediction method based on a deep reversible network according to claim 1 is characterized in that: The long-term and short-term integrated prediction module includes a plurality of long-term and short-term integrated prediction units, and the output feature of the current long-term and short-term integrated prediction unit is added to the input feature of the previous long-term and short-term integrated prediction unit to serve as the input feature of the next long-term and short-term integrated prediction unit; The long-term and short-term integrated prediction unit includes a short-term dependency block, a long-term and short-term aggregation block and an information update block; the prediction process of the short-term dependency block of the l-th long-term and short-term integrated prediction unit is expressed as: (3) (4) (5) (6) In the formula, , represents the trend gate and update gate of the short-term dependency block of the l-th long-term integrated prediction unit at time step t, It is the feature obtained by adding the spatiotemporal features predicted by the l-1th long-term and short-term integrated prediction unit at time step t and the input features of the l-2th long-term and short-term integrated prediction unit at time step t. is the spatiotemporal feature of the memory unit extracted by the l-1th long-term and short-term integrated prediction unit at time step t, , represents the weight matrix of the trend gate, , represents the weight matrix of the update gate, represents the hidden state features of the short-term dependency block of the l-th long-term integrated prediction unit at time step t, , The weight matrix representing the hidden state; represents the output features of the short-term dependency block of the l-th long-term and short-term integrated prediction unit at time step t, that is, the short-term spatiotemporal dependency features extracted by the l-th long-term and short-term integrated prediction unit at time step t; represents the convolution operation, Indicates activation operation. Indicates the number of frames contained in the time step; The prediction process of the long-term and short-term aggregation block of the l-th long-term and short-term integrated prediction unit is expressed as: (7) (8) In the formula, represents the long-term spatiotemporal dependency features obtained by the l-th long-term and short-term integrated prediction unit at time step t, represents the spatiotemporal characteristics of the memory units from the l-1-kth to the l-1th long-term and short-term integrated prediction units at time step t, represents the attention mechanism, represents the mixed spatiotemporal dependency features obtained by the l-th long-term and short-term integrated prediction unit at time step t; The update process of the information update block of the l-th long-term and short-term integrated prediction unit is expressed as: (9) (10) (11) (12) (13) In the formula, , represents the weight matrix and bias of the input gate, represents the input gate of the information update block of the l-th long-term and short-term integrated prediction unit at time step t, represents the fusion features of the information update block of the l-th long-term and short-term integrated prediction unit at time step t, , represents the weight matrix and bias of feature fusion, , represents the weight matrix and bias of the output gate, represents the output gate of the information update block of the l-th long-term and short-term integrated prediction unit at time step t, It is the spatiotemporal feature predicted by the lth long-term and short-term integrated prediction unit at time step t.

3. The spatiotemporal temperature prediction method based on a deep reversible network according to claim 1 or 2, characterized in that: The input image is transformed from 1×h×w to 4×h / 2×w / 2 through the pixel shuffling layer, and then the transformed image is split into two images of shape 2×h / 2×w / 2 as the input of the forward encoder; where h and w represent the height and width of the image, respectively.

4. The spatiotemporal temperature prediction method based on a deep reversible network according to claim 1, characterized in that: During the prediction model training and optimization process, the training loss is calculated according to the following objective optimization function; (16) In the formula, is the root mean square error, is the number of samples, , It is The true value and predicted value of each sample.

Images