CN-122019605-A - Atmospheric element downscaling prediction method based on deep learning

CN122019605ACN 122019605 ACN122019605 ACN 122019605ACN-122019605-A

Abstract

The invention discloses an atmospheric element downscaling prediction method based on deep learning, which particularly relates to the field of data processing of atmospheric element downscaling prediction, and comprises the steps of acquiring wind speed, temperature and humidity data by using a high-frequency sensor at an end side, performing compression processing and filtering processing on the acquired data, and generating a processed data stream; the invention constructs a time sequence calibration mechanism with time compensation, integrity screening, characteristic quantization and multistage feedback reasoning as cores in a terminal cloud cooperative link, so that the real evolution of terminal side situation perception and the cloud reasoning process keep the same time corresponding relation, and the problem of time phase dislocation caused by long-term asynchronization of a terminal side perception clock and a Yun Cetui-process clock is solved.

Inventors

XU CHENBO
HE YIN
SHI LEI
DENG ZIXIAO
LIU FULING
MA JUNLONG
CHEN ZHENGJIAN
Xiong Jinzhong
Chen Haiao
WANG YAXIN
Yu Shuran
XUE ZHIXIN
LI QIU
Liang Shanheng

Assignees

深能智慧能源科技有限公司
深能南京能源控股有限公司
深能北方能源控股有限公司

Dates

Publication Date: 20260512
Application Date: 20251226

Claims (8)

1. The atmospheric element downscaling prediction method based on deep learning is characterized by comprising the following steps of: S1, acquiring wind speed, temperature and humidity data by using a high-frequency sensor at an end side, performing compression processing and filtering processing on the acquired data to generate a processed data stream, transmitting the data stream to edge computing equipment, performing segmentation processing on the data stream by the edge computing equipment, and outputting processed data blocks divided according to time windows; S2, transmitting the processed data blocks to a cloud end through a communication link, calculating transmission delay for each data block by the cloud end, performing delay compensation on corresponding data block time stamps according to the calculated transmission delay, performing time synchronization operation based on the compensated time stamps, and outputting time-aligned data packets; s3, performing data integrity detection on the data packets, identifying the data packets meeting the integrity condition, and performing sorting processing on the data packets; calculating time offset, determining position index, analyzing wind speed, temperature and humidity change based on the sequencing result to generate local situation quantized values; s4, the cloud performs deep learning downscaling reasoning based on the feature set, the cloud invokes a time compensation algorithm to adjust a reasoning window according to the received end side timestamp, performs path optimization on data in the reasoning process, and outputs a downscaling prediction result; S5, the cloud end takes the downscaling prediction result as input, performs space reconstruction processing, generates a high-resolution weather prediction field, and transmits the high-resolution weather prediction field back to the end side through a communication link.
2. The atmospheric element downscaling prediction method based on deep learning of claim 1, wherein in S1, a high-frequency sensor is used at an end side to acquire wind speed data, temperature data and humidity data, and the acquired wind speed data, temperature data and humidity data are written into an end side data sequence according to an acquisition sequence; Performing compression processing on the end-side data sequence, calculating the difference value, the amplitude and the change direction of sampling points of the continuously acquired wind speed, temperature and humidity, recombining the calculated difference value, amplitude and change direction into a compressed data sequence according to time sequence, and outputting the compressed data sequence; Performing filtering processing on the compressed data sequence, and generating a filtered data stream by calculating the change rate of adjacent sampling points in the compressed data sequence, performing a rejection operation on sampling points with the change rate exceeding a preset upper limit, and simultaneously performing a smoothing operation on sampling points with the change rate being lower than a preset lower limit; Transmitting the filtered data stream to edge computing equipment, constructing time windows according to time information in the data stream by the edge computing equipment, dividing data in each time window into window segments according to time sequence, executing segmented extraction processing on the window segments by the edge computing equipment, extracting and recombining wind speed data, temperature data and humidity data in the window segments into processing data blocks according to time sequence, and outputting the processing data blocks.
3. The atmospheric element downscaling prediction method based on deep learning of claim 2, wherein in S2, the method further comprises transmitting the processed data blocks to a cloud end through a communication link, and recording the time stamp of each processed data block according to the received sequence at the cloud end to form a receiving sequence containing time stamp records; Calculating transmission delay differences for adjacent processing data blocks in the receiving sequence, attaching the transmission delay differences to corresponding processing data blocks to form a difference mark sequence, and judging each processing data block in the difference mark sequence: When the transmission delay difference value is in the threshold range, overlapping the transmission delay with the original time stamp to obtain a compensation time stamp; When the transmission delay difference exceeds a threshold range, marking the adjacent processed data blocks as abnormal data blocks, returning the abnormal marks to the corresponding positions of the received data sequences, and generating a reconstruction time stamp according to the sequence positions of the abnormal data blocks in the received data sequences; Writing the compensation time stamp and the reconstruction time stamp into corresponding processing data blocks respectively to form a compensation sequence, and performing time sequencing on the compensation sequence based on the compensation time stamp; In the sorting process, when the interval of the continuous compensation time stamp is detected to exceed the time synchronization threshold, writing the interval mark as a synchronization interval mark back to the compensation sequence, and readjusting the sorting result of the compensation sequence according to the synchronization interval mark; When the time synchronization threshold is not exceeded by the continuous compensation time stamp interval, the current compensation time stamp is used as a stable index to be written into a corresponding processing data block, the sequentially adjusted compensation sequences are sequentially combined in time sequence, and the time-aligned data packets are output.
4. The atmospheric element downscaling prediction method based on deep learning of claim 3 wherein S3 further comprises performing data integrity detection on the data packets, the data integrity detection specifically comprising sequentially reading wind speed data, temperature data, humidity data and time information for each data packet, and judging whether empty and unfilled items exist in the wind speed data, the temperature data and the humidity data; Writing the missing mark into the detection mark of the data packet when the missing item exists, writing the complete mark into the detection mark of the data packet when the missing item does not exist, and writing the data packet written with the detection mark into the data packet sequence according to the original sequence position; Sequentially selecting a current data packet and a previous data packet from the data packet sequence, reading time information of the current data packet and the previous data packet, and calculating a corresponding time interval; Writing a time continuous mark into the time mark of the current data packet when the time interval is between the upper limit of the preset time interval and the lower limit of the preset time interval, and writing a time abnormal mark into the time mark of the current data packet when the time interval exceeds the upper limit of the preset time interval or is lower than the lower limit of the preset time interval; And writing the data packet with the written time mark back to the position of the time mark in the data packet sequence, so that the data packet is overlapped with the time mark on the basis of the detection mark and is kept in the data packet sequence.
5. The atmospheric element downscaling prediction method based on deep learning of claim 4 wherein S3 further comprises reading wind speed data, temperature data and humidity data from a current data packet and a previous data packet selected in sequence in a data packet sequence, and calculating a wind speed difference value, a temperature difference value and a humidity difference value; When any difference value exceeds the preset change threshold range, writing a numerical value abnormal mark into the numerical value mark of the current data packet; Writing the current data packet written with the numerical value mark back to the position of the numerical value mark in the data packet sequence, so that the data packet is further overlapped with the numerical value mark on the basis of the detection mark and the time mark; Performing statistical processing on the detection mark, the time mark and the numerical value mark of each data packet in the data packet sequence, and writing the data packet into the complete data packet sequence when the detection mark is a complete mark, the time mark is a time continuous mark and the numerical value mark is a numerical value continuous mark; When the detection mark is a missing mark or the time mark is a time abnormal mark or the numerical mark is a numerical abnormal mark, the data packet is removed from the complete data packet sequence, so that the complete data packet sequence only contains the data packet meeting the integrity condition; The method comprises the steps of performing sorting processing on a complete data packet sequence according to time information to form a sorted data sequence, calculating time offset for each data packet in the sorted data sequence according to the sequence position of each data packet in the sorted data sequence, and recording the sequence position as a position index; and writing the time offset, the position index and the local situation quantized value into the feature set according to the sequence of the data packets and outputting the feature set.
6. The atmospheric element downscaling prediction method based on deep learning of claim 5 characterized by further comprising sequentially reading time offset, position index and local situation quantized value corresponding to each feature item from the feature set and generating an inference input record when each feature item is read in S4; Reading the end side timestamp of each inference input record in the inference input sequence, adding the end side timestamp and the time offset to generate a calculation time value, and writing the calculation time value into a calculation time field of the inference input record; Reading a calculated time value from each reasoning input record in the reasoning input sequence, and comparing the calculated time value with a preset upper compensation time limit and a preset lower compensation time limit: When the calculated time value is between the preset compensation time upper limit and the preset compensation time lower limit, writing the calculated time value into a calibration time field of the reasoning input record; When the calculated time value exceeds the upper limit of the preset compensation time or is lower than the lower limit of the preset compensation time, selecting a time field which is written into the inference input record and has smaller absolute value of the difference value between the calculated time value and the calculated time value from the upper limit of the preset compensation time and the lower limit of the preset compensation time, and writing the difference value between the time field and the end side timestamp into a time offset field of the inference input record; reading the calibration time field sequentially from the inference input sequence to form a calibration time sequence, selecting a time value at the initial position of the sequence as the initial time of an inference window according to the sequence position in the calibration time sequence, and selecting a time value at the end position of the sequence as the end time of the inference window; Performing difference calculation on the starting time and the ending time to generate a window span value, and comparing the window span value with a preset window span upper limit and a preset window span lower limit: and when the window span value exceeds the preset window span upper limit or is lower than the preset window span lower limit, generating a window abnormal mark and writing back the position of the corresponding reasoning input record in the reasoning input sequence to form a window feedback chain.
7. The atmospheric element downscaling prediction method based on deep learning of claim 6 characterized by further comprising the steps of sequentially reading a time offset field, a position index field and a local situation quantized value field in each inference input record for a window feedback chain, forming an inference input set by the time offset field, the position index field and the local situation quantized value field according to a reading sequence, inputting the inference input set into a deep learning downscaling inference model, and executing inference processing; In the reasoning process, sequentially reading intermediate reasoning values formed by each reasoning operation, generating a reasoning value sequence from the sequentially read intermediate reasoning values, performing difference calculation on adjacent reasoning values in the reasoning value sequence to generate a reasoning difference sequence, and comparing the reasoning difference sequence with a preset upper limit of a variation range and a preset lower limit of the variation range: generating a path adjustment mark when the reasoning difference value sequence is between the upper limit of the preset change range and the lower limit of the preset change range, writing the path adjustment mark back to the position corresponding to the reasoning input record in the reasoning input sequence to form a path adjustment chain, and re-executing deep learning scale-down reasoning based on the path adjustment chain; Sequentially reading the reasoning value sequences generated by re-executing the deep learning downscaling reasoning to generate a downscaling prediction result sequence, executing continuity difference calculation on adjacent reasoning values in the downscaling prediction result sequence to generate a continuity difference sequence, and comparing the continuity difference sequence with a preset continuity upper limit and a preset continuity lower limit: when the continuity difference value sequence is between a preset continuity upper limit and a preset continuity lower limit, generating a continuous output record; when the continuity difference value sequence exceeds the preset continuity upper limit or is lower than the preset continuity lower limit, an output adjustment mark is generated, the output adjustment mark is written back to the reasoning value sequence to form an output feedback chain, replacement processing is carried out on the output feedback chain, and the reasoning value after the replacement processing is used as a downscaling prediction result and is output.
8. The atmospheric element downscaling prediction method based on deep learning of claim 7 characterized by further comprising the steps of sequentially obtaining each inference value from the downscaling prediction result sequence to construct a spatial reconstruction input sequence, performing difference calculation on adjacent inference values in the spatial reconstruction input sequence to generate a spatial difference sequence, and solving the spatial difference sequence item by item for a local spatial variation; Comparing the local spatial variation with a preset upper variation limit and a preset lower variation limit: when any local spatial variation exceeds the preset variation upper limit or is lower than the preset variation lower limit, a spatial feedback chain is generated according to an inference value corresponding to the local spatial variation, and the spatial reconstruction input sequence is reconstructed based on the spatial feedback chain; Sequentially reading the time position and the numerical position of adjacent reasoning values for the reconstructed space reconstruction input sequence, performing proportion segmentation calculation on the adjacent reasoning values, dividing the time interval and the numerical interval between the two reasoning values into intermediate supplementary values in proportion, and generating an interpolation supplementary sequence arranged in time sequence; Based on the interpolation supplementary sequence, reading the spatial positions of adjacent reasoning values, performing difference calculation on the predicted values of the adjacent spatial positions, performing equidistant expansion operation on the spatial region covered by the interpolation supplementary sequence according to the difference value, and solving the supplementary predicted values of each position in the expansion region according to the distance ratio between the supplementary predicted values and the adjacent predicted values to form a spatial expansion sequence; Performing matrix difference calculation on predicted values of adjacent positions in the space reconstruction matrix to form a matrix difference sequence, and solving the matrix continuity quantity item by item for the matrix difference sequence; comparing the matrix continuity quantity with a preset continuity upper limit and a preset continuity lower limit: when the continuity quantity of each matrix is between a preset continuity upper limit and a preset continuity lower limit, combining the space reconstruction matrices to form a high-resolution weather prediction field; when any matrix continuity quantity exceeds a preset continuity upper limit or is lower than a preset continuity lower limit, generating a regional feedback chain according to a matrix position corresponding to the continuity quantity, executing reconstruction replacement processing based on the regional feedback chain, generating a high-resolution weather prediction field and outputting the high-resolution weather prediction field to the end side.

Description

Atmospheric element downscaling prediction method based on deep learning Technical Field The invention relates to the technical field of data processing of atmospheric element downscaling prediction, in particular to an atmospheric element downscaling prediction method based on deep learning. Background In an atmospheric element downscaling prediction system based on deep learning, a terminal side executes local situation sensing by a second-level rhythm by means of a high-frequency sensor, and captures fine-scale evolution of elements such as wind, temperature, humidity and the like in real time; However, in a real engineering scene, the end-side situation awareness link itself contains delays such as data acquisition, edge preprocessing, compression and uploading, and the cloud is also affected by factors such as network queuing, protocol encoding and decoding, message routing, task scheduling and reasoning resource preemption, so that an end-side clock and a Yun Cetui management clock are asynchronous for a long time at a physical level; the result is that the situation information uploaded by the end side characterizes the actual atmospheric evolution in the current or even near future, and the cloud actually used for inputting the downscaling reasoning still stays at the old situation several seconds or even tens of seconds before; even if cloud output is aligned to the same time stamp on a data format, the physical meaning of the cloud output is delayed by one or more evolution stages, and along with multi-period rolling operation of downscaling reasoning, the time deviation is accumulated and diffused gradually along a time sequence dependent chain of the model, so that high-resolution output is continuously deviated from the current situation of real atmosphere, and a recessive, accumulated and irreversible time phase dislocation error is formed; Therefore, in the existing end cloud coordination and situation awareness framework, the inherent asynchronism of an end-side awareness clock and a Yun Cetui management clock can not be effectively coordinated all the time, and the deep learning model is difficult to keep synchronous with continuous evolution of the real atmosphere. Disclosure of Invention In order to overcome the defects in the prior art, the embodiment of the invention provides an atmospheric element downscaling prediction method based on deep learning, which is used for constructing a time sequence calibration mechanism taking time compensation, integrity screening, feature quantization and multistage feedback reasoning as cores in an end cloud cooperative link, so that the real evolution of end-side situation awareness and a cloud reasoning process keep the same time corresponding relation, and the problem of time phase dislocation caused by long-term asynchronization of an end-side awareness clock and a Yun Cetui-stage clock in the background art is solved. In order to achieve the purpose, the invention provides the following technical scheme that the atmospheric element downscaling prediction method based on deep learning comprises the following steps: S1, acquiring wind speed, temperature and humidity data by using a high-frequency sensor at an end side, performing compression processing and filtering processing on the acquired data to generate a processed data stream, transmitting the data stream to edge computing equipment, performing segmentation processing on the data stream by the edge computing equipment, and outputting processed data blocks divided according to time windows; S2, transmitting the processed data blocks to a cloud end through a communication link, calculating transmission delay for each data block by the cloud end, performing delay compensation on corresponding data block time stamps according to the calculated transmission delay, performing time synchronization operation based on the compensated time stamps, and outputting time-aligned data packets; s3, performing data integrity detection on the data packets, identifying the data packets meeting the integrity condition, and performing sorting processing on the data packets; calculating time offset, determining position index, analyzing wind speed, temperature and humidity change based on the sequencing result to generate local situation quantized values; s4, the cloud performs deep learning downscaling reasoning based on the feature set, the cloud invokes a time compensation algorithm to adjust a reasoning window according to the received end side timestamp, performs path optimization on data in the reasoning process, and outputs a downscaling prediction result; S5, the cloud end takes the downscaling prediction result as input, performs space reconstruction processing, generates a high-resolution weather prediction field, and transmits the high-resolution weather prediction field back to the end side through a communication link. In a preferred embodiment, in S1, the method further includes acquiring wind speed