CN116128168A - Weather prediction method based on causal expansion convolution and Autoformer - Google Patents

Weather prediction method based on causal expansion convolution and Autoformer Download PDF

Info

Publication number
CN116128168A
CN116128168A CN202310404836.7A CN202310404836A CN116128168A CN 116128168 A CN116128168 A CN 116128168A CN 202310404836 A CN202310404836 A CN 202310404836A CN 116128168 A CN116128168 A CN 116128168A
Authority
CN
China
Prior art keywords
weather
prediction
module
network
trained
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310404836.7A
Other languages
Chinese (zh)
Inventor
荣欢
张译
蒋薇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Information Science and Technology
Original Assignee
Nanjing University of Information Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Information Science and Technology filed Critical Nanjing University of Information Science and Technology
Priority to CN202310404836.7A priority Critical patent/CN116128168A/en
Publication of CN116128168A publication Critical patent/CN116128168A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01W—METEOROLOGY
    • G01W1/00—Meteorology
    • G01W1/10—Devices for predicting weather conditions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Economics (AREA)
  • General Physics & Mathematics (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • Environmental & Geological Engineering (AREA)
  • General Health & Medical Sciences (AREA)
  • Marketing (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Business, Economics & Management (AREA)
  • Health & Medical Sciences (AREA)
  • Development Economics (AREA)
  • Operations Research (AREA)
  • Molecular Biology (AREA)
  • Quality & Reliability (AREA)
  • Primary Health Care (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Educational Administration (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Game Theory and Decision Science (AREA)
  • Atmospheric Sciences (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Ecology (AREA)
  • Environmental Sciences (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a weather prediction method based on causal expansion convolution and Autoformer, which is used for obtaining the prediction condition of a target future period about each weather phenomenon label according to each weather factor data collected in a target region preset history period and each weather phenomenon label, carrying out low-error weather prediction, and carrying out tuning on a network to be trained by constructing a causal expansion neural network module, constructing a weather prediction module based on Autoformer, constructing the network to be trained and a super-parameter selection module, and training the network to be trained to obtain a weather prediction model; and then, a weather forecast model is applied to forecast weather phenomena of the target area corresponding to the future time of the target. The invention has larger receptive field, can extract more information to learn the model prediction model, greatly improves the accuracy of model prediction and effectively reduces the model prediction error.

Description

Weather prediction method based on causal expansion convolution and Autoformer
Technical Field
The invention belongs to the field of natural science, and particularly relates to a weather prediction method based on causal expansion convolution and Autoformer, which aims to predict weather conditions when a certain amount of weather data are provided.
Background
Meteorological development is rapid in the 20 th century. Human knowledge of the atmospheric process is becoming more and more clear, as is the need for weather prediction. Weather forecast is an important means for weather work serving national economy and national defense construction. The weather forecast, especially the disaster weather forecast, is timely and accurately published by the weather desk through various channels, and plays an important role in protecting lives and properties of people, promoting economic development and the like.
The existing weather prediction method is divided into a traditional weather method and a numerical prediction method: the traditional weather method fills out meteorological data of the same level at the same time on a special graph, and the graph is called a weather graph. Through analyzing various meteorological elements on the weather map, a forecaster can know the distribution and structure of the current weather system (such as typhoons, fronts and the like) and judge the connection between the weather system and specific weather (such as rain, wind, fog and the like) and the future evolution condition of the weather system, so that weather forecast of various places is obtained. The time consumption is long, the requirements on manpower and material resources are large, and the prediction accuracy is not high; the weather forecast of the numerical forecasting method is calculated by a computer. Since the motion of the atmosphere follows some known laws of physics, according to these laws, the motion state of the atmosphere can be written into a set of partial differential equations, and as long as an initial value (the current condition of the atmosphere) is given, the variable value of the equation set changing with time can be solved, so that the future condition of the atmosphere can be obtained. The precision is high, the time consumption is less, and the consumption resources are less. Due to the development of the neural network in recent years, a time series prediction model based on the neural network has a great breakthrough in various weather prediction tasks, so that the application of the time prediction model to weather prediction is a research hotspot in the current weather field. However, the time prediction model based on the neural network has the defects in the parts of data processing, information feature extraction, neural network structure and the like, so that the weather prediction method based on the time prediction model is easy to generate prediction errors, and the prediction result is inaccurate.
Numerous researchers have made many searches to solve this problem. In the prior art, a scholars propose a double-synchronization period long-period and short-period memory cyclic neural network model, so that the prediction accuracy of periodic time sequence data is higher, and the error is smaller; a learner puts forward a double hidden layer recurrent neural network structure, and the fitting and representing capabilities of two groups of activation functions are utilized to provide the precision and generalization capability of the model; scholars also propose a multi-channel LSTM structure, which greatly reduces the number of parameters, reduces errors and improves prediction accuracy through a many-to-one structure; the above method reduces errors by optimizing the model structure. However, the error is still too large for the current weather prediction results, and the accuracy still needs to be improved. Therefore, it is necessary to use a new neural network model for weather prediction.
The causal expansion convolutional neural network structure has a larger receptive field, can convolve more data during each prediction, is convenient for feature extraction, is beneficial to model learning, and obtains more accurate prediction results.
Disclosure of Invention
Aiming at the problem that the weather result error is larger in calculation of the existing time prediction model, the invention provides a weather prediction method based on causal expansion convolution and Autoformer, which is used for realizing low-error weather prediction according to preset weather factor data acquired in a preset historical period of a target area and weather phenomenon labels, so as to obtain the prediction condition of the target future period under the preset weather phenomenon labels.
In order to achieve the above purpose, the present invention provides the following technical solutions:
the weather prediction method based on causal expansion convolution and Autoformer is characterized in that weather prediction models are obtained through steps S1 to S5 based on preset weather factor data and weather phenomenon labels under the condition that a target area corresponds to a preset historical time sequence, and weather phenomena of the target area corresponding to a target future time are predicted by applying the weather prediction models according to step i;
step S1: constructing a causal extended neural network module based on causal extended convolution, wherein the causal extended neural network module comprises causal convolution for processing time sequence problems and extended convolution for increasing model receptive fields, and then entering step S2;
step S2: constructing a weather prediction module based on an Autoformer, wherein the weather prediction module comprises a sequence decomposition unit, an encoder and a decoder; the sequence decomposition unit is used for smoothing the period item and highlighting the trend item; the encoder is used for gradually eliminating trend items to obtain period items; the decoder performs two-way processing, the upper branch processes the period item, the lower branch processes the trend item, and for the period item, the autocorrelation mechanism utilizes the period property of the sequence to aggregate subsequences with similar processes in different periods; for trend items, gradually extracting trend information from predicted hidden variables in an accumulated mode;
step S3: constructing a network to be trained, wherein the network to be trained comprises a data preprocessing module, a causal extended neural network module, a weather prediction module and an error learning module; the data preprocessing module is used for carrying out smoothing processing on input data; the input end of the data preprocessing module forms the input end of the network to be trained, the output end of the data preprocessing module is sequentially connected with the causal expansion neural network module, the weather prediction module and the error learning module in series, the output end of the error learning module forms the output end of the network to be trained, and then the step S4 is carried out;
step S4: the super parameter selection module performs tuning and optimizing on the network to be trained; the super-parameter selection module is used for properly allocating network parameters to be trained including the number of neurons, a loss function, an activation function and the batch size so as to obtain the optimal network performance;
step S5: each weather factor data and each weather phenomenon label are preset under the condition that the target area corresponds to a preset historical time sequence, each weather sample is obtained, each weather sample comprises the weather phenomenon label of a single time point corresponding to the target area, the preset time length a is arranged at intervals of the time point to the historical time direction, and each weather factor data is preset under the time sequence of continuous preset time length b; wherein a is greater than or equal to b;
based on each meteorological sample, taking the data of each preset meteorological factor in a time sequence in the meteorological sample as input, taking the weather phenomenon label in the meteorological sample as output, and training aiming at a network to be trained to obtain a trained model, namely a meteorological prediction model;
step i: and obtaining weather phenomenon labels of the target area corresponding to the target future time by obtaining the preset time length a of the target area corresponding to the target future time to the historical time direction and presetting each weather factor data under the time sequence of the continuous preset time length b and applying a weather prediction model.
Further, the data preprocessing module processes the data by adopting a mobile smoothing method.
Further, the super-parameter selection module adopts a particle swarm optimization algorithm to perform super-parameter optimization on the network to be trained, adjusts network parameters, optimizes the weight of the neural network, and performs feature selection and data normalization.
Further, the error learning module processes the linear fitting and the abnormal observed value to improve the accuracy of model prediction; the prediction error is estimated using the training error of the training samples for different situations.
Compared with the prior art, the weather prediction method based on causal expansion rolling and Autoformer has the following technical effects:
1. the causal expansion convolutional neural network module constructed by the invention utilizes a neural network structure with a larger receptive field and a multi-layer convolutional kernel to extract different information characteristics and gather the information characteristics into a characteristic diagram through multiple convolutions, thereby being beneficial to extracting more information to learn an image prediction model;
2. the weather prediction module is constructed based on an Autoformer, the sequence is decomposed to be used as an internal decomposition unit, the decomposition capacity of the complex time sequence is enhanced, and a similar periodic subsequence is polymerized by adopting an autocorrelation mechanism, so that the efficiency and the accuracy are better than those of a self-attention mechanism;
3. the error learning module constructed by the invention can greatly improve the accuracy of model prediction and effectively reduce the model prediction error.
Drawings
FIG. 1 is a flow chart of a causal extended convolution and Autoformer based weather prediction method;
FIG. 2 is a schematic diagram of a causal augmented neural network module;
FIG. 3 is a schematic diagram of an Autoformer-based weather prediction module;
FIG. 4 is a flow chart of the super parameter selection module.
Detailed Description
For a better understanding of the technical content of the present invention, specific examples are set forth below, along with the accompanying drawings.
Aspects of the invention are described herein with reference to the drawings, in which there are shown many illustrative embodiments. The embodiments of the present invention are not limited to the embodiments described in the drawings. It is to be understood that this invention is capable of being carried out by any of the various concepts and embodiments described above and as such described in detail below, since the disclosed concepts and embodiments are not limited to any implementation. Additionally, some aspects of the disclosure may be used alone or in any suitable combination with other aspects of the disclosure.
Referring to fig. 1 to 4, the causal extended convolution and auto former based weather prediction method according to the present invention will be described in further detail with reference to the embodiments. As shown in fig. 1, the causal extended convolution and Autoformer based weather prediction method comprises the following steps:
the first step: constructing a causal expansion neural network module based on causal expansion convolution;
and a second step of: constructing a weather prediction module based on an Autoformer;
and a third step of: constructing a network to be trained;
fourth step: the super parameter selection module adjusts the network parameters to be trained;
fifth step: training a network to be trained by using meteorological sample data to obtain a meteorological prediction model;
sixth step: and acquiring preset weather factor data, and predicting a corresponding weather phenomenon label by applying a weather prediction model.
In the example, firstly, a network to be trained is built, wherein the network to be trained comprises a data preprocessing module, a causal expansion neural network module, a weather prediction module and an error learning module; the data preprocessing module processes data by adopting a mobile smoothing method; the causality expansion neural network module increases the receptive field of the network by using expansion convolution, and more data can be convolved during each prediction to obtain more accurate prediction results.
As shown in fig. 2, in the causal extended neural network module, one-dimensional weather time series data is input respectively, different features are extracted by using a plurality of convolution kernels, and then the features are gathered into a feature map through a plurality of convolutions, wherein the output convolution operation of each layer can be described as follows:
Figure SMS_1
；
in the method, in the process of the invention,
Figure SMS_2
is the firstuOutput of layer->
Figure SMS_3
Is a restricted linear activation unit,/->
Figure SMS_4
Is the firstsThe bias term of the individual feature map is,Ais the total number of layers for the firstsThe characteristic map of the image is shown in the figure,Yis kernel size, ++>
Figure SMS_5
Is a characteristic diagramsWeight of->
Figure SMS_6
Is the firstsLayer pair one-dimensional weather time series data->
Figure SMS_7
Is the first of (2)u:u+Y-1Data.
The convolution operation results of all the layers are input into a collecting layer for maximum selection, and the maximum value in the collecting layer is selected and output as follows:
Figure SMS_8
；
wherein:Tis the span of the span,Ois the size of the pool layer and,qis a one-dimensional single-layer vector,ris the receptive field size;
the output of the sink layer is then processed using the ReLU activation function, the output being:
Figure SMS_9
；
wherein, the liquid crystal display device comprises a liquid crystal display device,Ais a learnable parameter;
Figure SMS_10
and the output of the causal expansion neural network module is input into a weather prediction module.
The weather prediction module adopts an Autoformer depth decomposition architecture, and comprises an internal sequence decomposition unit, an autocorrelation mechanism and a coder-decoder; autoformer decomposes the sequence as an internal unit, embedded in the codec; in the prediction process, autoformer alternately optimizes the prediction result and sequence decomposition, and gradually separates trend items and period items from input data.
The sequence decomposition unit smoothes the period term and the prominent trend term based on the idea of moving average, and the sequence decomposition process is summarized as follows:
Figure SMS_11
；
the specific process is as follows:
Figure SMS_12
；
Figure SMS_13
；
wherein, the liquid crystal display device comprises a liquid crystal display device,
Figure SMS_14
is a hidden variable to be decomposed, < ->
Figure SMS_15
A trend term is indicated and a trend term is indicated,/>
Figure SMS_16
representing a periodic term.
As shown in FIG. 3, the encoder pair in the weather prediction module inputs time series data
Figure SMS_17
Performing a series of operations such as autocorrelation, sequence decomposition, forward transmission, etc., and eliminating the weather trend component +.>
Figure SMS_18
，/>
Figure SMS_19
Only the seasonal component of the weather is retained +.>
Figure SMS_20
，/>
Figure SMS_21
；And transmitting the seasonal component to the decoder for helping the decoder to improve the prediction result; for the total number of layers isNEncoder of->
Figure SMS_22
The individual encoder layers are summarized as: />
Figure SMS_23
；
The specific process is as follows:
Figure SMS_24
；
wherein, the liquid crystal display device comprises a liquid crystal display device,
Figure SMS_25
；/>
Figure SMS_28
indicate->
Figure SMS_31
The output of the individual encoder layers,/>
Figure SMS_27
From->
Figure SMS_30
Embedding; />
Figure SMS_32
Indicate->
Figure SMS_34
The>
Figure SMS_26
Weather seasonal components of the layer; />
Figure SMS_29
Indicate->
Figure SMS_33
The>
Figure SMS_35
Meteorological trend component of the layer.
The decoder is divided into two-way processing modes, the upper branch processes weather seasonal components, the lower branch processes weather trend components, and the time series data is input
Figure SMS_36
Obtaining input weather seasonal component->
Figure SMS_37
And the trend component of the meteorological pattern->
Figure SMS_38
The process of (2) is as follows:
Figure SMS_39
；
wherein:
Figure SMS_40
，/>
Figure SMS_41
representing a zero padding placeholder and +.>
Figure SMS_42
Average value of (2).
As shown in fig. 3, in the process of processing the weather seasonal component by the upper branch, extracting the time dependence in the prediction state by using the autocorrelation layer 1, then extracting information from the smooth weather seasonal sequence output by the encoder by using the autocorrelation layer 2, and finally entering the forward transmission layer; the total layer number of the decoder is N, the th
Figure SMS_43
The individual decoders are summarized as: />
Figure SMS_44
；
The specific process is as follows:
Figure SMS_45
；
wherein, the liquid crystal display device comprises a liquid crystal display device,
Figure SMS_48
；/>
Figure SMS_50
indicate->
Figure SMS_52
The output of the decoder layer,/>
Figure SMS_47
From->
Figure SMS_54
Obtaining;
Figure SMS_55
indicate->
Figure SMS_56
The>
Figure SMS_46
Weather seasonal components of the layer; />
Figure SMS_49
Indicate->
Figure SMS_51
The>
Figure SMS_53
Meteorological trend component of the layer.
The lower branch processes the meteorological trend component, and each decoder adopts weighted addition to separate each sub-layer of the upper branch
Figure SMS_57
Together, the weather trend component of the decoder is obtained by the following steps:
Figure SMS_58
；
wherein, the liquid crystal display device comprises a liquid crystal display device,
Figure SMS_59
from->
Figure SMS_60
Obtaining; />
Figure SMS_61
Indicate->
Figure SMS_62
Each layer of meteorological trend components extracted from each decoder
Figure SMS_63
Projection to +.>
Figure SMS_64
Is a weight matrix of (a).
The error learning module improves the accuracy of model prediction by linear fitting and abnormal observation value processing, and the specific process is as follows:
weather prediction based module
Figure SMS_65
Training result set obtained by secondary training and sequenced from small to large
Figure SMS_66
And training error set corresponding to each training result +.>
Figure SMS_67
The result of the prediction is obtained as follows>
Figure SMS_68
Corresponding prediction error->
Figure SMS_69
：
Case 1: if present
Figure SMS_70
Make->
Figure SMS_71
Get +.>
Figure SMS_72
The corresponding training error serves as a prediction error:
Figure SMS_73
；
case 2: if there is a pair of
Figure SMS_74
Make->
Figure SMS_75
Then the prediction error is obtained by linear fitting: />
Figure SMS_76
；
Case 3: if it is
Figure SMS_77
Taking the average value of all negative errors as the predictionError: />
Figure SMS_78
；/>
Wherein the method comprises the steps of
Figure SMS_79
Representing training error set->
Figure SMS_80
All negative errors in>
Figure SMS_81
Representing the number of negative errors;
case 4: if it is
Figure SMS_82
Taking the average value of all positive errors as the prediction error: />
Figure SMS_83
；
Wherein the method comprises the steps of
Figure SMS_84
Representing training error set->
Figure SMS_85
All positive errors in>
Figure SMS_86
Indicating the number of positive errors;
obtaining and predicting results
Figure SMS_87
Corresponding prediction error->
Figure SMS_88
Then, the prediction output is: />
Figure SMS_89
。
After the establishment of the network to be trained is completed, adjusting and optimizing the parameters of the network to be trained by using the super parameter selection module; the super-parameter selection module adopts a particle swarm optimization algorithm to allocate the quantity, the time, the loss function and the activation function of neurons of a network to be trained and the batch size; as shown in fig. 4, the number of neurons, the optimizer, the epoch, the loss function, the activation function, the batch size are particles in the algorithm, and form a particle swarm, when the particle swarm algorithm runs an update iteration, the particle position and the particle velocity are updated, and the fitness value is calculated, and the particle swarm algorithm comprises the following calculation processes:
Figure SMS_90
；
where d is the particle position, v is the particle velocity, y represents the degree of importance of d and v, p represents the optimal particle population derived from the history of all populations, g is the inertial factor, h is the individual learning factor, r is a random number from 0 to 1;
and calculating to obtain an optimal particle swarm p, namely, the optimal value of the parameters of the network to be trained, namely, the number of neurons, an optimizer, an epoch, a loss function, an activation function and the batch size, and completing the parameter tuning of the training network.
Then, training the network to be trained by using the sample data; each group of sample data respectively forms a length of from Monday to Friday of the week and Saturday weather phenomena
Figure SMS_91
Weather sample data->
Figure SMS_92
The method comprises the steps of carrying out a first treatment on the surface of the The meteorological sample data of each group form meteorological time series data +.>
Figure SMS_93
The method comprises the steps of carrying out a first treatment on the surface of the Wherein, the meteorological factors comprise temperature, humidity, air pressure, cloud cover and wind power grade; weather phenomena include sunny days, cloudy days, overcast days, light rain, medium rain, heavy rain and thunder gust; repeatedly training the network to be trained by taking the meteorological factor data from Monday to Friday in each group of meteorological sample data as input and the weather phenomenon of Saturday as output; and obtaining a trained model, namely a weather prediction model.
While the invention has been described in terms of preferred embodiments, it is not intended to be limiting. Those skilled in the art will appreciate that various modifications and adaptations can be made without departing from the spirit and scope of the present invention. Accordingly, the scope of the invention is defined by the appended claims.

Claims (4)

1. The weather prediction method based on the causal expansion convolution and the Autoformer is used for obtaining the prediction condition of the target future period with respect to each weather phenomenon label in the preset history period according to the preset weather factor data and each weather phenomenon label acquired in the target region preset history period, and carrying out weather prediction; the method is characterized in that weather phenomenon prediction models are obtained through steps S1 to S5 based on the preset weather factor data and the weather phenomenon labels of the target area corresponding to the preset historical time sequence, and weather phenomena of the target area corresponding to the target future time are predicted by applying the weather prediction models according to step i;
step S1: constructing a causal extended neural network module based on causal extended convolution, wherein the causal extended neural network module comprises causal convolution for processing time sequence problems and extended convolution for increasing model receptive fields, and then entering step S2;
step S2: constructing a weather prediction module based on an Autoformer, wherein the weather prediction module comprises a sequence decomposition unit, an encoder and a decoder; the sequence decomposition unit is used for smoothing the period item and highlighting the trend item; the encoder is used for gradually eliminating trend items to obtain period items; the decoder performs two-way processing, the upper branch processes the period item, the lower branch processes the trend item, and for the period item, the autocorrelation mechanism utilizes the period property of the sequence to aggregate subsequences with similar processes in different periods; for trend items, gradually extracting trend information from predicted hidden variables in an accumulated mode;
step S3: constructing a network to be trained, wherein the network to be trained comprises a data preprocessing module, a causal extended neural network module, a weather prediction module and an error learning module; the data preprocessing module is used for carrying out smoothing processing on input data; the input end of the data preprocessing module forms the input end of the network to be trained, the output end of the data preprocessing module is sequentially connected with the causal expansion neural network module, the weather prediction module and the error learning module in series, the output end of the error learning module forms the output end of the network to be trained, and then the step S4 is carried out;
step S4: the super parameter selection module performs tuning and optimizing on the network to be trained; the super-parameter selection module is used for properly allocating network parameters to be trained including the number of neurons, a loss function, an activation function and the batch size so as to obtain the optimal network performance;
step S5: each weather factor data and each weather phenomenon label are preset under the condition that the target area corresponds to a preset historical time sequence, each weather sample is obtained, each weather sample comprises the weather phenomenon label of a single time point corresponding to the target area, the preset time length a is arranged at intervals of the time point to the historical time direction, and each weather factor data is preset under the time sequence of continuous preset time length b; wherein a is greater than or equal to b;
based on each meteorological sample, taking the data of each preset meteorological factor in a time sequence in the meteorological sample as input, taking the weather phenomenon label in the meteorological sample as output, and training aiming at a network to be trained to obtain a trained model, namely a meteorological prediction model;
step i: and obtaining weather phenomenon labels of the target area corresponding to the target future time by obtaining the preset time length a of the target area corresponding to the target future time to the historical time direction and presetting each weather factor data under the time sequence of the continuous preset time length b and applying a weather prediction model.
2. The causal extended convolution and auto former based weather prediction method according to claim 1, wherein the data preprocessing module processes data using a mobile smoothing method.
3. The causal expansion convolution and Autoformer-based weather prediction method according to claim 1, wherein the super parameter selection module performs super parameter optimization on a network to be trained by adopting a particle swarm optimization algorithm, adjusts network parameters, optimizes neural network weights, and performs feature selection and data normalization.
4. The causal extended convolution and auto former based meteorological prediction method of claim 1, wherein the error learning module improves accuracy of model prediction with linear fitting and outlier processing; the prediction error is estimated using the training error of the training samples for different situations.
CN202310404836.7A 2023-04-17 2023-04-17 Weather prediction method based on causal expansion convolution and Autoformer Pending CN116128168A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310404836.7A CN116128168A (en) 2023-04-17 2023-04-17 Weather prediction method based on causal expansion convolution and Autoformer

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310404836.7A CN116128168A (en) 2023-04-17 2023-04-17 Weather prediction method based on causal expansion convolution and Autoformer

Publications (1)

Publication Number Publication Date
CN116128168A true CN116128168A (en) 2023-05-16

Family

ID=86303129

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310404836.7A Pending CN116128168A (en) 2023-04-17 2023-04-17 Weather prediction method based on causal expansion convolution and Autoformer

Country Status (1)

Country Link
CN (1) CN116128168A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117235445A (en) * 2023-11-09 2023-12-15 广东工业大学 Weather prediction system and method based on multiple models and time sequences

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115080964A (en) * 2022-08-16 2022-09-20 杭州比智科技有限公司 Data flow abnormity detection method and system based on deep learning of graph
CN115453051A (en) * 2022-08-24 2022-12-09 同济人工智能研究院(苏州)有限公司 Gas concentration prediction and real-time positioning method for gas monitoring equipment
CN115759474A (en) * 2022-12-08 2023-03-07 北京九栖科技有限责任公司 Multitask load prediction method and system based on Autoformer model and XGboost model
CN115758641A (en) * 2022-11-22 2023-03-07 贵州电网有限责任公司 Power transmission line icing prediction method based on auto-former-progressive decomposition model
CN115829139A (en) * 2022-12-13 2023-03-21 同济人工智能研究院(苏州)有限公司 Unstable power load prediction method and system based on time series analysis model

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115080964A (en) * 2022-08-16 2022-09-20 杭州比智科技有限公司 Data flow abnormity detection method and system based on deep learning of graph
CN115453051A (en) * 2022-08-24 2022-12-09 同济人工智能研究院(苏州)有限公司 Gas concentration prediction and real-time positioning method for gas monitoring equipment
CN115758641A (en) * 2022-11-22 2023-03-07 贵州电网有限责任公司 Power transmission line icing prediction method based on auto-former-progressive decomposition model
CN115759474A (en) * 2022-12-08 2023-03-07 北京九栖科技有限责任公司 Multitask load prediction method and system based on Autoformer model and XGboost model
CN115829139A (en) * 2022-12-13 2023-03-21 同济人工智能研究院(苏州)有限公司 Unstable power load prediction method and system based on time series analysis model

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117235445A (en) * 2023-11-09 2023-12-15 广东工业大学 Weather prediction system and method based on multiple models and time sequences
CN117235445B (en) * 2023-11-09 2024-01-26 广东工业大学 Weather prediction system and method based on multiple models and time sequences

Similar Documents

Publication Publication Date Title
Ghaderi et al. Deep forecast: Deep learning-based spatio-temporal forecasting
CN114626512B (en) High-temperature disaster forecasting method based on directed graph neural network
Wu et al. Evolving RBF neural networks for rainfall prediction using hybrid particle swarm optimization and genetic algorithm
CN113128793A (en) Photovoltaic power combination prediction method and system based on multi-source data fusion
CN112766549A (en) Air pollutant concentration forecasting method and device and storage medium
CN111079980B (en) Optical power prediction method based on self-adaptive classification strategy and hybrid optimization SVR
CN109165693A (en) It is a kind of to sentence knowledge method automatically suitable for dew, frost and the weather phenomenon of icing
CN113705877A (en) Real-time monthly runoff forecasting method based on deep learning model
Liu et al. Heating load forecasting for combined heat and power plants via strand-based LSTM
CN113469427A (en) Convolution LSTM-based photovoltaic power station day-ahead power generation amount prediction method and system
CN116128168A (en) Weather prediction method based on causal expansion convolution and Autoformer
WO2022039675A1 (en) Method and apparatus for forecasting weather, electronic device and storage medium thereof
CN116415730A (en) Fusion self-attention mechanism time-space deep learning model for predicting water level
CN117374956A (en) Short-term prediction method for photovoltaic power generation of comprehensive energy station
CN116384583A (en) Photovoltaic power prediction method based on multiple neural networks
CN114970946A (en) PM2.5 pollution concentration long-term space prediction method based on deep learning model and empirical mode decomposition coupling
CN116341391B (en) Precipitation prediction method based on STPM-XGBoost model
CN113111419A (en) Method and system for establishing and predicting air-conditioning load prediction model in office building
CN116485010B (en) S2S precipitation prediction method based on cyclic neural network
CN110766219A (en) Haze prediction method based on BP neural network
Sharma et al. Deep Learning Based Prediction Of Weather Using Hybrid_stacked Bi-Long Short Term Memory
CN115481788A (en) Load prediction method and system for phase change energy storage system
CN114648095A (en) Air quality concentration inversion method based on deep learning
KR20230066927A (en) Method and apparatus for predicting power generation
Wang et al. The prediction method of tropical cyclone intensity change based on deep learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20230516