CN117574960A

CN117574960A - Multi-factor plum rainfall prediction method based on self-adaptive graph structure and reinforcement integration

Info

Publication number: CN117574960A
Application number: CN202311718382.7A
Authority: CN
Inventors: 陈圣劼; 王禹; 桑小卓; 朱志伟
Original assignee: Jiang Sushengqixiangtai
Current assignee: Jiang Sushengqixiangtai
Priority date: 2023-12-14
Filing date: 2023-12-14
Publication date: 2024-02-20

Abstract

The invention provides a multi-factor plum rainfall prediction method based on self-adaptive graph structure and reinforcement integration, which comprises the following steps: collecting historical data of factors related to plum rain, normalizing, mining association relations among a plurality of factors, and constructing a multi-factor adjacency matrixAnd constructing a plum rain amount prediction model based on a graph convolution and time convolution neural network by adopting the adjacency matrix, training after constructing, and training by using MSE as a loss return model until the model reaches fitting, wherein the calculation formula of the MSE is as follows:wherein,is a predicted value, Y _i Is the true tag value. The prediction method can effectively capture and represent the timeliness characteristics of plum rain data and the change trend of the timeliness characteristics under the influence of various factors.

Description

Multi-factor plum rainfall prediction method based on self-adaptive graph structure and reinforcement integration

Technical Field

The invention relates to the field of plum rain prediction, in particular to a multi-factor plum rain prediction method based on self-adaptive graph structure and reinforcement integration.

Background

In the eastern Asia monsoon area of China, a concentrated rainfall period with the characteristic of plum rain weather is always reserved in river areas of river and yam in 6-7 months each year, the rainy days of the plum rain period are concentrated, the rainfall is large, the heavy rain is frequent, and the duration of the juvenile parts is long, so that a large-scale flood disaster is often caused in the downstream area in the Yangtze river. The plum rain is a unique weather and climate phenomenon in the middle and downstream regions of the Yangtze river of China, and is a stage of the east Asia subtropical monsoon rain belt pushing to the river basin.

Plum rain is an important characteristic of plum rain, and besides short-time flood caused by the influence of a hot band cyclone, the total trend of drought and flood in the flood season in the middle and downstream regions of the Yangtze river is mainly determined by the abundance apology of plum rain. The late and early arrival of plum rain and the quantity of plum rain in the middle and lower reaches of the Yangtze river directly influence summer-harvest summer seeds, flood prevention, flood control, drought resistance and water storage in the regions. At present, a lot of forecasting work surrounds the forecasting of the rainfall in the plum flood period, and at present, experience and statistical forecasting methods are widely applied to the forecasting of the plum rain. Statistical methods include univariate time series methods, multivariate statistical methods, and the like. Overall, the single variable time series method has low accuracy in forecasting the rainfall in the plum flood season, which may be related to the influence of the continuous change of the relevant physical factors as the forecasting factors, and has great limitation on the continuous change of the self evolution rule. Multivariate statistical methods such as regression prediction models have evolved to some extent in business prediction and scientific research. Studies have revealed that the effects of plum rain anomalies are affected and controlled by multiple physical processes and influencing factors, including the sea temperature anomalies in the Pacific, indian and Atlantic, the Tibet plateau and European asiatic distribution, the external forcing signals of North-south North sea ice and soil humidity, etc., and the atmospheric circulation factors of the Western Pacific subtropical high pressure, the south Asia high pressure, the Western wind rapid flow, the European high latitude land activities, etc. (Huang Ronghui, etc., 2003; liang Ping, etc., 2007). Abnormal amount of plum rain per year is often the result of multi-factor synergy, and complex interactions and links exist between different influencing factors. Peng Duan et al (2005) predict the precipitation in the 6 th county (city) season of Zhaoqing by applying stepwise regression equation, and select several principal components of the northern hemisphere 500hPa and the Pacific ocean thermal field as factors (1748 in total). The factors comprise main information of a climate background field, and a stepwise regression model is established by selecting the factor with the largest correlation coefficient through the correlation screening of the predicted quantity and the predicted factor group. The result shows that the established forecast equation has high complex correlation coefficient and good forecast effect on precipitation in the onset flood period. Liu Dandan (2011) predicts the plum rain amount in Yiwu city by using a BP neural network model, and finds that the BP neural network model has certain advantages compared with the traditional stepwise regression model. However, the climate system itself has high nonlinearity, factors influencing the plum rainfall are not independent of each other, and the traditional statistical forecasting model does not consider nonlinear interaction between forecasting factors, so that the prediction of the plum rainfall is still greatly limited.

In view of this, the present invention has been made.

Disclosure of Invention

In view of the above, the invention discloses a multi-factor plum rain amount prediction method based on self-adaptive graph structure and reinforcement integration, which can effectively capture and represent the timeliness characteristics and the change trend of plum rain data under the influence of various factors. According to the method, a graph structure is used for modeling plum rain data affected by various factors, correlation among different variables is captured through a graph convolution network, meanwhile, a prediction model is respectively built through three deep neural networks of a gate control loop unit (GRU), a Time Convolution Network (TCN) and a Transformer, and the three deep networks are combined through a reinforcement learning method, so that adaptability and robustness of the model are effectively improved.

Specifically, the invention is realized by the following technical scheme:

the invention provides a multi-factor plum rainfall prediction method based on self-adaptive graph structure and reinforcement integration, which comprises the following steps:

collecting historical data of factors related to plum rain, normalizing, mining association relations among a plurality of factors, and constructing a multi-factor adjacency matrix

And constructing a plum rain amount prediction model based on a graph convolution and time convolution neural network by adopting the adjacency matrix, training after constructing, and training by using MSE as a loss return model until the model reaches fitting, wherein the calculation formula of the MSE is as follows:

Wherein,is a predicted value, Y _i Is the true tag value.

The invention researches the plum rain amount because the plum rain amount is an important characteristic of plum rain, and besides short-time flood caused by the influence of a hot zone cyclone, the general trend of drought and waterlogging in the flood season in the middle and downstream regions of the Yangtze river is mainly determined by the abundance apology of the plum rain amount. The late and early arrival of plum rain and the quantity of plum rain in the middle and lower reaches of the Yangtze river directly influence summer-harvest summer seeds, flood prevention, flood control, drought resistance and water storage in the regions. At present, a lot of forecasting work surrounds the forecasting of the rainfall in the plum flood period, and at present, experience and statistical forecasting methods are widely applied to the forecasting of the plum rain. Statistical methods include univariate time series methods, multivariate statistical methods, and the like. Overall, the single variable time series method has low accuracy in forecasting the rainfall in the plum flood season, which may be related to the influence of the continuous change of the relevant physical factors as the forecasting factors, and has great limitation on the continuous change of the self evolution rule. Multivariate statistical methods such as regression prediction models have evolved to some extent in business prediction and scientific research. Studies have revealed that the effects of plum rain anomalies are affected and controlled by multiple physical processes and influencing factors, including the sea temperature anomalies in the Pacific, indian and Atlantic, the Tibet plateau and European asiatic distribution, the external forcing signals of North-south North sea ice and soil humidity, etc., and the atmospheric circulation factors of the Western Pacific subtropical high pressure, the south Asia high pressure, the Western wind rapid flow, the European high latitude land activities, etc. (Huang Ronghui, etc., 2003; liang Ping, etc., 2007). Abnormal amount of plum rain per year is often the result of multi-factor synergy, and complex interactions and links exist between different influencing factors. Peng Duan et al (2005) predict the precipitation in the 6 th county (city) season of Zhaoqing by applying stepwise regression equation, and select several principal components of the northern hemisphere 500hPa and the Pacific ocean thermal field as factors (1748 in total). The factors comprise main information of a climate background field, and a stepwise regression model is established by selecting the factor with the largest correlation coefficient through the correlation screening of the predicted quantity and the predicted factor group. The result shows that the established forecast equation has high complex correlation coefficient and good forecast effect on precipitation in the onset flood period. Liu Dandan (2011) predicts the plum rain amount in Yiwu city by using a BP neural network model, and finds that the BP neural network model has certain advantages compared with the traditional stepwise regression model. However, the climate system itself has high nonlinearity, factors influencing the plum rainfall are not independent of each other, and the traditional statistical forecasting model does not consider nonlinear interaction between forecasting factors, so that the prediction of the plum rainfall is still greatly limited.

Physical factors affecting plum rain mainly include:

(1) Quaternary wind circulation factor

Atmospheric flow anomalies are a direct cause of precipitation anomalies. Therefore, the atmospheric circulation change in the plum rainy period not only directly determines the distribution of rain belts in the east Asia area in the plum rainy period, but also determines the intensity of plum rain. Existing studies on precipitation anomalies indicate that some important atmospheric circulation systems have an important effect on precipitation in eastern asia. In the circulation systems, the west Pacific subtropical high pressure has a very close relation with the east Asia climate abnormality, the west Pacific subtropical high pressure ridge line is located in the south, and the southwest warm and humid air flow at the lower part of the eastern troposphere of China replaces the dry southeast air flow, thereby creating favorable water vapor conditions for the dewatering in the Yangtze river basin of China. Zhang Zihan et al (2008) analyze the characteristics of the circulating background and the strong precipitation in the rainy period of Zhejiang plum in 2008 to obtain the western stretching ridge point and the north boundary position of the western Pacific side height, and have important indication significance for the occurrence of the strong precipitation in the rainy period of Zhejiang plum. Further, studies on plum rain by Bao Yuanyuan et al (2009) have found that the northern pacific equatorial band affects the variation of the secondary height and leads to the variation of the secondary height, and has important indication significance for forecasting the coming and going of plum rain from Jianghuai.

The area, intensity, west extension ridge point and ridge index definitions (Liu Yunyun, etc., 2012) for west tai pair are as follows:

the secondary high area index is the sum of areas surrounded by lattice points with the potential heights of not less than 5880 potential meters on the potential height field of 500hPa within the range of 110 DEG E-180 DEG E below 10 DEG N.

The secondary high intensity index is the cumulative value obtained by subtracting 5870 potential meters from the potential height value of 500hPa greater than 5870 potential meters lattice point in the above range.

The paragao-west ridge point is the longitude of the most west lattice point of 5880 position potential meters in the range of 90 DEG E-180 DEG E at the north of 10 DEG N. If the angle is 90 degrees E, uniformly counting to be 90 degrees E; if there is no 5880 potential meter contour for a month, the maximum value of the years history for that month is substituted.

The secondary high ridge index is the average value of the latitude position of the latitudinal wind shear line in the secondary tropical high-pressure body surrounded by the 5880 potential meter contour line within the range of 110 DEG E-150 DEG E from 10 DEG N (i.e. u=0,)。

south Asia high pressure is a powerful and stable semi-permanent high pressure circulation at the upper layer of the upper convection layer and the lower layer of the lower convection layer in the south Asia in summer, and is an important member in the atmospheric high-rise subtropical high pressure system. Zhang Qiong et al (2000) analyzed the effect of a 100hPa high field on the precipitation in summer of our country by SVD method, found that when the 100hPa high field is abnormally strong, the river basin is abnormally rainy. When the south Asia high pressure is located above the Qinghai-Tibet plateau and is in the eastern part, the strength is stronger, and the subtropical high pressure is stronger, thereby being beneficial to river-to-river precipitation; the south Asia high pressure is strong, and the deepening of the polar vortex is also beneficial to the river and the river. The area index of the south Asia high voltage is the total lattice point number of the 100hPa isobaric surface in summer, 30 DEG E is recorded as the intensity index of the south Asia high voltage by the total lattice point number of the potential height in the southeast Asia high voltage area which is not less than 16760 potential meters, and the total sum of the lattice point value of the 100hPa potential height which is more than 16760 potential meters and the difference between 16760 potential meters. The east-extension ridge point of south asia high pressure is represented by a longitude value of 16760 potential meter contour east ridge point (Hu Jinggao et al, 2010; tu Houwang et al, 2020).

In addition, the interaction among the quaternary wind subsystems in the Asia-Pacific quaternary wind region has very significant influence on river basin precipitation. Ding Yihui and Liu Yunyun (2008) found that there was a distant relationship between the precipitation of the monsoon in the southwestern india and the precipitation of the rainy period in chinese plum. Meanwhile, the plum rain in the Yangtze river basin is also affected by the summer monsoon of the North Pacific, and is in inverse relation with the precipitation of the summer monsoon of the North Pacific, namely when the summer monsoon of the North Pacific weakens, the precipitation of the Yangtze river basin in summer is too much. Dai Xingang et al (2002) found that indian convective precipitation can cause a change in the high-pressure locations of the western pacific subtropical zone to affect east asian summer monument and chinese summer precipitation. The indian summer season wind index is located at 5 ° to 15 ° N,40 ° to 80 ° E and 20 ° to 30 ° N, 850hPa weft pitch level differences in the 70 ° to 90 ° E region (Wang et al, 2001). The north pacific summer monsoon index is located at a difference between the 850hPa weft wind gap levels in the regions 5 ° to 15 ° N,100 ° to 130 ° E and 20 ° to 30 ° N,110 ° to 140 ° E (Wang and Fan, 1999).

TABLE 1 Quaternary wind circulation factor index

(2) Other circulation factors

Because the precipitation is generated under the interaction of the systems with various scales, the precipitation in the period of the river and the river plum rain on the circular flow field is also subjected to the multi-scale interaction of each member in the monsoon system. For the study of the rainfall in summer and particularly the plum rain in the eastern part of China, not only the influence of medium-low weft circulation such as the tropical monsoon system and the subtropical monsoon circulation needs to be considered, but also the influence of medium-high weft circulation and related cold air activities.

The invasion of dry and cold air from middle and high latitudes and middle and high floors is an important motive and thermal cause for the formation and maintenance of a plum rain humidity front. The passage of cold air under the south of asia in high weft-blocking situations has an effect on abnormal precipitation in the rainy days of the plum (Zhang Qingyun and Tao Shiyan, 1998;Chen and Zhai,2014;Park and Ahn,2014). Zhang Qingyun and Tao Shiyan (1998) found that abnormal precipitation in the summer plum rain phase of east asia is closely related to medium and high latitude blocking, and when high pressure is established and stabilized in the jaw Huo Cike sea, a "+ - +" type of wave train is easily formed from the flat field in the middle and high latitude of asia and the eastern region of east asia, and the distribution situation and interaction of the wave train from the flat region of eastern asia often cause more precipitation in the summer of east asia, particularly in the plum rain phase; conversely, when the jaw Huo Cike sea is in a low value region, the high latitude in Asia and the distance-to-flat field in eastern Asia are prone to form "- + -" distance-to-flat wave trains, and the distribution situation of the eastern Asia distance-to-flat wave trains is unfavorable for precipitation in the plum rainy period in eastern Asia summer. Yang Yiwen (2001) proposes that the 7 month east Asia occlusion situation is an important circulation system that causes excessive rain floods in the Yangtze river basin of China, with historically severe Yangtze river years (e.g., 1954 and 1998) often accompanied by the occurrence of the 7 month east Asia occlusion situation. Jin Ronghua et al (2008) found that Ula mountain blocking high pressure is important in 2007 in flood storms in river basins. The blocking high voltage index adopts a one-dimensional blocking index (TM index for short) proposed by Tibaldi and Molteni (1990), and the blocking event is defined by judging that the upper potential height of a certain longitude of a 500hPa altitude field is reversed, and the index is used for daily monitoring business of a national climate center. The basic principle is as follows:

Where Z is the potential height (gpm),latitude,>is 3 reference latitudes> Delta= -5 °, 0 °, 5 °. For at least one delta value for a certain longitude at a certain time, if the condition is satisfied: GHGS > 0 and GHGN- < 10, i.e., confirm that the longitude has a blocking high pressure at that time (taking the day in the middle of 5 days as that time). The blocking high pressure index used is the maximum value of each longitude GHGS, and the larger the index is, the stronger the blocking high pressure is.

In addition, east Asia subtropical high altitude western wind rushing is one of the important systems affecting east Asia, and its position shift and intensity changes are closely related to east Asia atmospheric flow season adjustment and plum rain volume (Ni Yunqi and Zhou Xiu, 2004; kuang Xueyuan and Zhang Yaocun, 2006). Kuang Xueyuan and Zhang Yaocun (2006) when the torroid in the east Asia subtropical western wind is abnormal and is north-oriented, the torroid in the river basin is weak, the lower layer is scattered abnormally, the ascending air flow is weak, the precipitation in the plum rain season is little, otherwise, the precipitation in the year of the torroid is more in the south of the torroid abnormality, the torroid is strong, and flood disasters are easy to happen. Ding Yihui et al (2007) have further studied the temporal and spatial characteristics of plum rains and have found that a high altitude western wind rapid flow is an evacuation flow region in the southerly region of the plum rains and a rising movement region on the right side of the eastern wind rapid flow inlet region of the tropics, which is advantageous for the occurrence, development and maintenance of plum rains. Jin Ronghua et al (2012) define the east Asia subtropical western wind rapid-flow intensity index as the average of 200hPa latitudes over each longitude in the critical zone (30-37.5 DEG N, 110-130 DEG E), and define the east Asia subtropical western wind rapid-flow position index as the average of the latitudes over each longitude in the critical zone (20-55 DEG N, 110-130 DEG E) where the maximum of 200hPa latitudes over each longitude is located.

North-polar surge (AO) is a north-south teeter-totter oscillation of the barometric field in arctic and mid-latitude areas. AO and its changes are closely related to the atmospheric circulation in the northern hemisphere and to climate change in many places. Li Chongyin et al (2008) analysis finds that the abnormality of the AO in March causes stronger northeast wind in the middle and lower reaches of Yangtze river, and the wind field in the lower layer of the troposphere has obvious anti-cyclone circulation above the Yangtze river, and the radiation situation of the low-layer wind field of the troposphere is not favorable for the confluence of water vapor in the region, so that the precipitation of plum rain is less. The AO index definition is from Climate Prediction Center (CPC, U.S. climate prediction center) and is defined as the temporal coefficient of the first modality of the EOF analysis of sea level barometric pressure from the level in areas outside the northern hemisphere of winter, https:// www.esrl.noaa.gov/psd/data/climatendics/list/.

Meanwhile, antarctic surge (AAO) is taken as the most main mode of high-latitude atmospheric flow in the southern hemisphere, reflects the inversion phase change characteristic of the air pressure field between the subsidiary tropical high-pressure zone and the high-latitude extremely-low-pressure zone of the southern hemisphere, and can influence the rainfall in the eastern summer of the later period. Bao Xuejun et al (2006) found that the late 6-7 months Yangtze river basin precipitation would increase when the current period is 4-5 months AAO is stronger, and vice versa. High brightness et al (2003) also indicate that in spring, especially when 5 months have abnormal strong movement in the antarctic wave, the method often corresponds to more precipitation in river and river areas in summer, later plum blossom, and longer plum period; conversely, when the movement of the antarctic waves in the current period is extremely weak, the river basin is prone to less precipitation. AAO index definitions are likewise from CPC, https:// www.esrl.noaa.gov/psd/data/claateinendices/list +.

Therefore, the prediction of summer precipitation at the middle and lower reaches of the Yangtze river should comprehensively consider the changes of each influencing factor of high, medium and low latitude atmospheric circulation.

TABLE 2 other circulation factor indices

(3) Tropical sea Wen Yinzi

The China is in the eastern Asia monsoon area, the sea Liu Reli difference is used as a first driving force, the thermal characteristics of the sea determine that the sea plays an extremely important role in climate change, so that among a plurality of factors influencing the rainfall in summer of China, the sea temperature factor is particularly important, wherein the sea surface temperature abnormality of the Pacific ocean and the tropical Indian ocean has a remarkable influence on the climate of China. The ENSO is a main mode of the large-scale sea-air interaction of the tropical sea, influences atmospheric circulation through various remote correlations, is the strongest signal for causing annual scale climate abnormality, and is also a key factor for Asian monsoon abnormality and the occurrence of drought and waterlogging in China. In the next summer, which occurs in early Nino (Lannia), the convective activity around the Philippines is stronger, the North Pacific side tropical high pressure is stronger (weaker), resulting in more (less) precipitation in the Yangtze river basin and North of the Jiangnan, and less (more) precipitation in the Jianghuai river basin (Jin Zuhui and Tao Shiyan, 1999;Huang et al,1989; chen Wen, 2002). Further research indicates that the influence of el nino on the precipitation in the flood season of China depends on the phase. Shi Jiuen et al (1983) indicated that in the early days of el nino, the flood season in the Yangtze river basin was more watershed and the north and south precipitation was less watershed; in the following early elnnino, the Yangtze river basin is slightly lower in precipitation in the flood season and more in the north and south. Huang Ronghui et al (2003) consider that in the early stage of development, the summer precipitation is more in river basin and less in yellow river basin and North China; in the early stage of attenuation, the precipitation situation is opposite to that of the early stage, the river basin is less, and the river basin is more in the south of the Yangtze river. The scholars further discussed the effect of el nino on the next year of summer precipitation in combination with different types of el nino, different end times, different decay types, etc. Li and Ni Yunqi (1997) indicate that early in the next summer, the middle and downstream Yangtze river is heavily watered. Yuan Yuan et al (2012) classified the ENSO to find that the next year in which eastern el Nino occurs, precipitation in the Jianghai region exhibits a north-south difference, particularly in the Yangtze river basin and the majority of the north thereof, while precipitation in the Jiangxi and Hunan regions of the Yangtze river is much less. The middle part of the Yankee river is controlled by southwest wind distance level in the middle and downstream regions of the Yangtze river in summer of the middle part of the Ei Nino, and is controlled by an abnormal anti-cyclone in the southeast coast of China, so that the river basin is an intersection region of a low-layer wind field, which is favorable for converging water vapor of the Western Pacific ocean in the river basin, thereby being favorable for more precipitation in the river basin of China; meanwhile, the western Pacific auxiliary tropical high pressure is obviously stronger to the west and more obvious than the auxiliary high condition corresponding to the eastern type, so that the main rain belt is also more north than the eastern type, and is mainly distributed in the river basin between the Yangtze river and the yellow river. The mixed type early Nino next year in summer yellow river basin and north and south China have more precipitation, and river regions have less precipitation. The maximum sea Wen Zhengju level is mainly distributed in the eastern Pacific ocean of the equator and near the coast of the south america when the el Nino event is developed to the full period, namely Nino3 zone (5 DEG N-5 DEG S,150 DEG W-90 DEG W) is defined as eastern el Nino; the maximum sea Wen Zhengju level is mainly distributed near the pacific boundary line in the equator when the el Nino develops to the full period, namely a Nino4 zone (5 DEG N-5 DEG S,160 DEG E-150 DEG W), and is classified as a middle-type el Nino; the distribution of the maximum sea Wen Zhengju level between the two in the early stages of progression of el nino is essentially in the nino3.4 region (5°n-5°s,170°w-120°w), defined as a mixed el nino. Defining the average autumn and winter average Sea Surface Temperature Abnormality (SSTA) of 3 key areas, determining different types of early Nino events, and using the early Nino events as a predictor of summer plum rain. Chen Sheng and He Jinhai (2017) further distinguish that different attenuation modes of different el Nino types have different effects on the rainfall in summer of China, and indicate that the possibility of drought bias between two rivers in summer and in the upstream region of the eastern el Nino year is obviously increased when E-W type is attenuated, and the south China is slightly more. W-E eastern type early Nino in the next year, the precipitation is more in the two rivers and upstream areas, and the large-value center of abnormal precipitation is mainly located in the areas along the river. For the 3 attenuation modes of the middle-part el nino, the abnormal distribution of the precipitation in China in summer in the next year also shows obvious differences: the precipitation abnormal large value zone is mainly positioned between the yellow river and the Huai river during S-shaped attenuation, possibly appears in the Yangtze river basin during the P-shaped attenuation stage, and is mainly positioned in the region downstream of the yellow river during the S-shaped next year; the S and A type of the next year are mostly northeast, especially northeast, precipitation is little, while the P type of the next year is mostly northeast, precipitation is obviously much; in southwest areas, the S-type next year precipitation is more general, while the A-type next year areas are easy to be more drought; in northwest regions, the A-type next-year precipitation is unusual, while the S-type and P-type next-year precipitation is unusual.

Chen Lieting et al (1985) found that although the tropical ocean temperature anomaly was not as strong as the equatorial east Pacific ocean, its latitudinal thermodynamic contrast strength was maintained as flat as the equatorial east Pacific ocean temperature, and that the Indian ocean was also an important factor affecting the rain level of the east Asia plum. There are two main modes of tropical indian ocean temperature abnormalities: sea temperature uniform mode (Indian Ocean Basin-Wide, IOBW) and indian ocean dipole type oscillation (IOD) throughout the indian ocean.

IOBW is the most dominant modality of change in sea temperature in the indian ocean of the tropics, defined as the sea temperature in the indian ocean region of the tropics (20 ° S-20 ° N,40 ° to 100 ° E), usually starting to develop in winter, reaching its strongest in the next spring. The peak summer of el nino in the next year, the spring and summer tropical indian ocean warm sea temperature anomaly may excite tropical atmospheric kelvin waves and induce the ackman irradiance mechanism in the north pacific to cause the occurrence of north pacific anti-cyclone, thereby causing the eastern summer monsoon precipitation anomaly including the mermaid rain, i.e., the "capacitor" effect of tropical indian ocean in the effect of ENSO on eastern asia climate. In addition, ocean temperatures in india can occur and persist significantly anomaly-indian ocean dipole type oscillations (IOD) (Saji et al, 1999), which can alter atmospheric flows through processes such as remote correlation to affect summer precipitation in china. The analysis of the dipole index and 160 sites of China in summer precipitation shows that the IOD and the China in summer precipitation are better in 6-8 months. The average SSTA of the tropical west ocean (10 ° S to 10 ° N,50 ° to 70 ° E) region minus the average SSTA of the tropical east ocean (10 ° S to 10 ° N,90 ° to 110 ° E) region was taken as the IOD index according to the method of Saji (1999). In recent years, more and more scholars consider that a certain relation exists between the IOD and the ENSO, li and the like (2006) find that the IOD increases along with the water vapor flux and the vertical wind field of the Yangtze river basin when the ENSO occurs, and the cyclone is abnormal. Thus, a composite index IODN3 index is defined, the IODN3 index being the sum of the Nino3 index and the IOD index, i.e. IODN 3=ssta [ (5 ° N to 5 ° S)

150-90 DEG W) ] +SSTA [ (10 DEG S-10 DEG N, 50-70 DEG E) ] -SSTA [ (10 DEG S-10 DEG N, 90-110 DEG E) ]. In the new index, the Nino3 index and the IOD index have equivalent amplitude, and especially the new comprehensive index can better represent the occurrence of the strong el Nino event and the strong Indian ocean dipole event, and the influence on the summer atmosphere and the rainfall of China is reasonable when the two events occur simultaneously by using the index.

The western pacific warm pool area is also a sea temperature critical area affecting the plum rain of the river and the river, and the critical influence period is 12 months of the first 1 year to 1 month of the current year (early winter), the sea temperature of the western pacific warm pool area is abnormally higher in the winter of the current period, the plum rain amount is abnormally higher in most areas of the river and the river in the same year, and vice versa (Mao document and the like, 2007). The average sea surface temperature of the key area (8-18 DEG N, 128-146 DEG E) of the Western Pacific heating pool in the early winter is selected to represent the abnormality of the Western Pacific heating pool.

Zong Haifeng et al (2005) explored the relationship between the rainy days of the Yangtze river basin and the global sea temperature, found that the relationship between the plum rain and the sea temperature in the early winter is the most close, and the main factors affecting the precipitation in the plum rain period are different in different time scales. The main factors affecting quasi-2 year oscillations are the black tide sea temperature of the subtropical zone, the eastern Pacific sea temperature of the equator, the Indian sea temperature of the eastern Pacific sea of the equator and the Atlantic subtropical sea temperature of the south, all contributing uniformly. The 2 main factors affecting the 3-5 year oscillation are substantially the same as the region of the 2 factor of the quasi-2 year oscillation. That is, there are variations in sea temperature in these 2 regions, both in the ENSO cycle and in the quasi-2 year oscillation. The major factors affecting quasi-10 year oscillations and the internationally varying differences occur in temperate sea areas, the major factors being North Pacific ocean drift sea temperature and Alaska ocean current sea temperature and south Atlantic temperate sea temperature. The main factors of the linear trend component are the eastern ocean sea temperature of the equator, the ocean current sea temperature of California, the ocean sea temperature of the equator, the ocean black tide sea temperature of the North Pacific ocean, the ocean drift sea temperature of the west wind, the ocean sea temperature of the temperate zone Atlantic sea temperature and the like.

(4) Other seas Wen Yinzi

The rain amount of Jianghuai plum is influenced by tropical sea and also by middle and high latitude sea (Lin Jian and He Jinhai, 2000). The early winter and spring north atlantic sea temperature bias heating can possibly excite European Asia wave columns to lead the Ula mountain and the jaw Huo Cike sea to be in blocking high pressure enhancement, so that the winter and the spring north atlantic sea temperature bias heating can lead the winter and the spring north atlantic sea temperature bias heating to be early in plum entering and the precipitation increase. In addition, the relationship between the eastern ocean and the plum rain in the North Pacific ocean in the subtropical zone is most remarkable on the annual scale, and the plum rain precipitation is influenced by influencing the change of the low-layer wind field and the sea temperature in the North Pacific ocean in the tropical zone. When the sea temperature annual signals of the North Pacific ocean in the subtropical zone are in the normal phase, the low-layer air pressure in the Western Pacific ocean in the equatorial region is higher, and western wind is prevalent in the east Pacific ocean in the tropical zone, the depth of the east Pacific thermocline in the tropical zone is increased due to the abnormal western wind, the sea temperature on the surface layer is higher, and the early Nino event is easy to occur, so that the precipitation of the Mei rain is more. The average sea temperature of 8 points [ (12°n,160°e), (14°n,170°e), (16°n,180 °), (18°n,170°w), (20°n,160°w), (22°n,150°w), (24°n,140°w), (30°n,130°w) ] in the eastern region of the auxiliary tropical north pacific from 7 months of the first 1 year to 6 months of the current year is defined as auxiliary tropical north pacific sea Wen Yinzi, reflecting the change in the auxiliary tropical north pacific sea temperature.

The north atlantic sea temperature can drive the atmospheric circulation change through evaporation and atmospheric processes, thereby affecting the northern hemisphere atmospheric planet scale circulation and thus the eastern asia climate. North atlantic sea Wen Sanji is the most prominent modality of north atlantic sea annual change rate, and the tripolar distribution, in which the sea temperature range from north to south appears as "- +" ("+ - +"), is defined as north atlantic sea Wen Sanji (negative) index. Gu et al (2009) found that early winter North Atlantic sea Wen Sanji in the positive (negative) phase, and that summer plum rain in the middle and downstream of the Yangtze river was more (less) in precipitation. The North Atlantic sea Wen Sanji adopts a national climate center standard, and takes a North Atlantic (0-60 DEG N, 80-0 DEG W) sea-temperature range flat field (linear trend is removed) EOF first mode as a projection mode, and the sea-temperature range flat field of a real-time month is projected onto the mode after the global sea-temperature warming influence is removed, so as to obtain the North Atlantic sea Wen Sanji sub-index.

TABLE 3 sea temperature factor index

(5) Snow and sea ice factor

Ice and snow circle is one of the important members of the climate system, and its changes can have an impact on the global climate. A series of studies have indicated that plateau or European snow cover plays an important role in the change of atmospheric circulation and the formation of precipitation in the eastern region of China (Charney and Shukla,1981; chen Xingfang and Song Wenling, 2000; huang,1985; yan Kailin et al, 2019).

Researches show that Qinghai-Tibet Gao Yuandong and spring snow are abnormal for years, and the north of Yangtze river and Jiangnan in summer is prone to more precipitation; the asiatic teenagers are rainless in the river basin in summer, rainy in North China and south China (Chen Xingfang and Song Wenling, 2000). Meanwhile, when the surface heat flux of Qinghai-Tibet plateau is abnormally high in spring, the river and the river are rich in plum rain. In spring and summer, the Qinghai-Tibet plateau is used as an atmospheric heat source, the near-ground layer is a thermal low pressure, when the surface heat flux of the plateau is enhanced, the thermal low pressure of the ground causes the air flow to be combined to generate large-scale convection activity, and the surface heat is brought to a high layer, so that the whole convection layer above the air is warmer, the 200hPa south Asia high pressure is stronger, and meanwhile, the auxiliary high west stretching is stronger. The southwest rapid flow with strong north side of the auxiliary high brings rich water vapor to be collected in the regions from Jianghuai to east China, so that stronger potential is unstable, and the water vapor is beneficial to abnormal deviation of precipitation in the plum rain period in the river basin (Yan Kailin and the like, 2019) corresponding to the strong rising centers of the river region and the sky above the east China.

In addition to Qinghai-Tibet plateau, research has found that North Asia (60-135E, 32-75N, north Asia for short) continents in winter have a significant effect on Jianghuai plum rain in the flood season of China (Chen Sheng, etc., 2012). When the surface heat-sensing cold source of the northland Asia in winter in the current period is stronger, the cold air strength at the north side of the plum rain front in summer is greatly enhanced, the northeast cold vortex is enhanced, and the transportation of the cold air to the south is enhanced; the position of the western wind rapid flow of the auxiliary tropical zone at the high altitude of 7 months is south, the south Asia high pressure is of the Qinghai-Tibet high pressure type, the rising movement of the river basin is enhanced, the upper layer of the river basin is dry, the lower layer of the river basin is wet, the instability of the atmospheric layer junction is increased, the convection activity is enhanced, and the precipitation of plum rain is facilitated. When the earth surface heat-sensing cold source is weaker in winter, the situation is opposite. The intensity of the north asian continental winter surface-sensing heat sink was characterized by defining a normalized moment plane of average surface-sensing heat flux of the eastern siberian plain, the vicinity of the barkarsh lake, and the northeast 3 major areas of China (centered at 86°e,58°n, 79°e,49°n and 120°e,49°n, 25 lattice points each around) as winter north asian continental heat-sensing source intensity index WSH.

The change of sea ice in the key areas of winter and spring excites an atmospheric abnormal Rossby wave source (Honda et al, 2009) through abnormal turbulent heat flux, and the atmospheric energy fluctuation propagates to east Asia in the form of wave trains, so as to influence the east Asia circulation in summer and further influence river and river precipitation (Zhang Renan, etc. 2018). When the Loidejun et al (2022) finds that the north pole sea ice of the former winter is abnormally much through regression and related analysis, the north cold air and the southwest warm-wet air flow at the high-pressure periphery of the subtropical zone are intersected at the middle and downstream of the Yangtze river, so that the region has abnormal wind field and water vapor radiation and is extremely much in precipitation. And a multi-regression prediction model of the plum rain amount in the middle and lower reaches of the Yangtze river is constructed by utilizing sea temperature and sea ice factors, and the fitting and prediction effects are good. Zhang Renan et al (2018) found that north and Bafen Bay in early spring had little ice (more) and north Bafen had ice (more) in summer based on CAM3.1 numerical simulations, while in the Yangtze river basin area of our country there was significantly less precipitation. The opposite is true when there is too much sea ice in the northern and Bafen Bay regions of the balun sea. The method utilizes the spring sea ice index and the late spring dipole type snow index to establish a prediction model of the summer precipitation of the river and river basin, and the return result shows that the river and river basin summer precipitation prediction model has higher prediction skills on the annual change rate of the summer precipitation of the river and river basin.

(6) Solar activity

Solar activity is generally referred to as moving bodies in localized areas of the sun, which often accumulate in areas called active areas, including blacks, spots, spectrum spots, daily lugs, flare, and coronal mass casts, along with rapid changes in time. Solar activity strongly affects the earth's space environment, the earth's climate, the magnetic field, the ionosphere, the earth's atmosphere, etc. Numerous studies have shown that weather and climate change on the timescale of hours to decades and centuries is significantly affected by solar activity. Li et al (2002) studied the time-space evolution law of summer precipitation in the eastern region 1880-1999 of China, and found that changes in summer precipitation in the middle and downstream of the Yangtze river, the Huaihe river basin and the North China region are significantly positively correlated with solar activity on the chronology time scale (including the penultimate time scale).

TABLE 4 snow, sea ice and solar activity factor

In view of the above effects on the amount of plum rain, the present invention provides a method for predicting the amount of plum rain according to local conditions, which comprises collecting and preprocessing historical data of a plurality of factors related to plum rain, modeling the plum rain data under the influence of a plurality of factors by using a graph structure, capturing correlations between different variables by using a graph convolution network, simultaneously, respectively building prediction models by using three deep neural networks of a gate control loop unit (GRU), a Time Convolution Network (TCN) and a transform, and combining the three deep networks by using a reinforcement learning method. Compared with the traditional integration method, the reinforcement learning method is an algorithm with continuous intelligent trend through interaction between the intelligent agent and the environment, so that the reinforcement learning method obtains better effect in the integration and optimization process. In addition, the integration of multiple deep learning methods can effectively improve the adaptability and robustness of the model. The method can effectively capture and represent the timeliness characteristic and the change trend of the plum rain data under the influence of various factors.

Drawings

Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments, wherein the accompanying drawings are only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the invention, and wherein like reference numerals represent like parts throughout the several views, and wherein:

FIG. 1 is a schematic diagram of a specific flow structure of a prediction method according to the present invention;

FIG. 2 is a schematic diagram of the structure of three models of the present invention;

FIG. 3 is a reinforcement learning schematic diagram of the present invention;

FIG. 4 is a flow chart of the training method of the present invention;

fig. 5 is a schematic flow chart of a computer device according to an embodiment of the present invention.

Detailed Description

Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings, wherein like reference numerals in the various drawings indicate identical or similar elements unless otherwise indicated, and wherein the implementations described in the following exemplary embodiments do not represent all implementations consistent with the present disclosure, but rather are merely examples of apparatuses and methods consistent with some aspects of the present disclosure as detailed in the accompanying claims,

The terms used in this disclosure are used solely for the purpose of describing particular embodiments and are not intended to be limiting of the disclosure, as used in the present disclosure and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise, it being understood that the term "and/or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items,

it should be understood that although the terms first, second, third, etc. may be used in this disclosure to describe various information, these information should not be limited to these terms, which are merely used to distinguish one type of information from another, e.g., a first information may also be referred to as a second information, and similarly a second information may also be referred to as a first information, depending on context, as the term "if" as used herein may be interpreted as "at … …" or "at … …" or "in response to a determination".

Example 1

The embodiment of the invention adopts N related to plum rain for the previous T daysTime series data of factorsAnd an adjacency matrix A representing the association relation among the N factors is used as the input of the time sequence convolution network, and the plum rain amount Y on the T+1th day is predicted.

Time series data of N factors related to plum rain in past T daysAnd an adjacency matrix->To describe the correlation between these N factors as input to the time series convolution network to predict the amount Y of plum rain on day t+1. By combining time dynamics with complex correlations between factors, the network aims to improve the accuracy of plum rain predictions. The method specifically comprises the following steps, and a flow diagram of the whole method is shown in fig. 1:

step 1: historical data of a plurality of factors related to plum rain is collected and preprocessed, in particular as follows:

step 1-1: historical data of factors related to plum rain, the former as training samples, and Mei Yuliang corresponding to time, the latter as training labels, were collected. The present invention relates to 29 factors, namely n=29, and the specific table 5 below shows:

TABLE 5 factor species

Step 1-2: normalizing the historical data of the factors by adopting a formula (1);

step 2: excavating a plurality ofThe association relation between factors, and then a multi-factor adjacency matrix is constructedThe method comprises the following steps:

step 2-1: constructing two groups of adaptively adjustable factor weight matrixesAnd->Wherein c represents the dimension of the parameter that can be learned in the weight matrix;

Step 2-2: calculating an adjacency matrix by adopting a formula (2) for constructing an adaptive graph structure;

step 3: building a plum rain amount prediction model based on graph convolution and time convolution neural network, wherein the model comprises four layers:

(1) The first layer is input layer, and the input data is preprocessed data setAnd an adjacency matrix A;

(2) The second layer is a graph roll-up neural network layer, and the data setAnd performing two-layer graph rolling operation with the adjacent matrix A, wherein the specific implementation is shown in a formula (3):

wherein σ (·) and ReLU (·) are activation functions, W ₀ And W is ₁ Is the weight of the two-layer picture volume lamination.

(3) The third layer is a prediction network layer, which comprises three kinds of neural networks including a gate control loop unit (GRU), a Time Convolution Network (TCN) and a transducer, and the three kinds of models are schematically shown in FIG. 2. GRU and transducer are 1 layer; since the historical time step is 12, the prediction step is 1, the time convolution network is set to be 5 layers, the convolution kernels are 3 in size, the convolution kernels are 16, 32, 32, 40 and 40 respectively, the expansion coefficients are 1,2,4,4,2 respectively, and finally a convolution layer with the size of 1×1 convolution kernel is added to convert the output dimension of the time convolution network into 151×1×1.

(4) The fourth layer is a network integration layer, and three learnable weights w are set ₁ ，w ₂ ，w ₃ To weight sum the output of the prediction network layer to obtain the final prediction result, wherein w ₁ Is the weight of GRU, w ₂ Is the weight of TCN, w is the weight of transducer.

Step 4: training the learnable weight of the network integrated layer by a reinforcement learning method, setting the attenuation coefficient lambda to be 0.5, the learning rate gamma to be 0.95 and the maximum iteration number to be 50, as shown in fig. 3, specifically comprising the following steps:

step 4-1: a state matrix S and an action matrix a are constructed. The state matrix is a weight of three deep networks in the network integration layer. The action matrix is an action of weight adjustment.

S＝[w ₁ ,w ₂ ,w ₃ ]#(12)

a＝[Δw ₁ ,Δw ₂ ,Δw ₃ ]#(13)

Step 4-2: a loss function L, a bonus function R and an evaluation function Q are constructed.

Step 4-3: and training the agent according to the training set. According to the current state S, the agent executes action a.

Step 4-4: and calculating a loss function L to obtain a reward R, and formulating a next strategy.

Wherein A (i) isThe actual measurement data in the training set is used to determine,is predictive data in the training set.

Step 4-5: and calculating an evaluation function Q and updating a Q table.

Q _m+1 (S _m ,a _m )＝Q _m (S _m ,a _m )+

γ _m (R(S _m ,a _m )+λmaxQ _m (S _m+1 ,a _m+1 )-Q _m (S _m ,a _m ))#(16)

Where λ is the attenuation coefficient and γ is the learning rate.

Step 4-6: repeating steps 4-3 to 4-5 until the iteration stop condition is satisfied. I.e. the state matrix S is the optimal weight for the three current deep networks.

Step 5: training an overall model: and (3) inputting a training sample into the neural network in the step (3), comparing the output of the model with the label data obtained in the step (2), and returning to the model to train by using MSE as a loss until the model is fitted. The formula for the MSE is as follows:

wherein,is a predicted value, Y _i Is the true tag value.

Step 6: and (3) training is completed, a final model is obtained, and the flow of the training method is shown in fig. 4.

According to the embodiment of the invention, the time series data is modeled by adopting the self-adaptive graph structure, and the corresponding self-adaptive graph structure is constructed aiming at priori knowledge under different conditions, so that the real-time characteristic change of the data in the space dimension is better captured and reflected. And respectively establishing a prediction model by using three deep neural networks of GRU, TCN and transducer, and combining the three deep neural networks by using a reinforcement learning method, so that the adaptability and the robustness of the model are effectively improved.

Experimental example 1

After comparing the above example 1 with other prediction methods in the prior art, specific results are shown in table 6 below:

table 6 comparison of predicted results

Wherein,

from the above predictions in table 6, it can be seen that the RMSE of the prediction scheme of the present invention is lower than the MAE compared to other schemes, which means that the prediction of the amount of plum rain is more accurate.

The invention provides a method for predicting plum rain amount, and also provides a system for predicting the plum rain amount, which comprises the following steps:

and a diagram construction module: the historical data for collecting factors related to plum rain is normalized, and the association relation among a plurality of factors is mined, so that a multi-factor adjacency matrix is constructed/>

The prediction model building module: the method is used for constructing a plum rain amount prediction model based on a graph convolution and time convolution neural network by adopting the adjacency matrix, training is carried out after construction, the MSE is used as a loss return model for training until the model is fitted, and the calculation formula of the MSE is as follows:

wherein,is a predicted value, Y _i Is the true tag value.

In the implementation, each module may be implemented as an independent entity, or may be combined arbitrarily, and implemented as the same entity or several entities, and the implementation of each unit may be referred to the foregoing method embodiment, which is not described herein again.

Fig. 5 is a schematic structural diagram of a computer device according to the present disclosure. Referring to FIG. 5, the computer device 400 includes at least a memory 402 and a processor 401; the memory 402 is connected to the processor via a communication bus 403, and is configured to store computer instructions executable by the processor 401, and the processor 401 is configured to read the computer instructions from the memory 402 to implement the steps of the prediction method according to any one of the embodiments.

For the above-described device embodiments, reference is made to the description of the method embodiments for the relevant points, since they essentially correspond to the method embodiments. The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the objectives of the disclosed solution. Those of ordinary skill in the art will understand and implement the present invention without undue burden.

Computer readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media and memory devices including, for example, semiconductor memory devices (e.g., EPROM, EEPROM, and flash memory devices), magnetic disks (e.g., internal magnetic disks or removable disks), magneto-optical disks, and CD-ROM and DVD-ROM disks. The processor and the memory can be supplemented by, or incorporated in, special purpose logic circuitry.

Finally, it should be noted that: while this specification contains many specific implementation details, these should not be construed as limitations on the scope of any invention or of what may be claimed, but rather as descriptions of features of specific embodiments of particular inventions. Certain features that are described in this specification in the context of separate embodiments can also be implemented in combination in a single embodiment. On the other hand, the various features described in the individual embodiments may also be implemented separately in the various embodiments or in any suitable subcombination. Furthermore, although features may be acting in certain combinations and even initially claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.

Similarly, although operations are depicted in the drawings in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. In some cases, multitasking and parallel processing may be advantageous. Moreover, the separation of various system modules and components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

Thus, particular embodiments of the subject matter have been described. Other embodiments are within the scope of the following claims. In some cases, the actions recited in the claims can be performed in a different order and still achieve desirable results. Furthermore, the processes depicted in the accompanying drawings are not necessarily required to be in the particular order shown, or sequential order, to achieve desirable results. In some implementations, multitasking and parallel processing may be advantageous.

The foregoing description of the preferred embodiments of the present disclosure is not intended to limit the disclosure, but rather to cover all modifications, equivalents, improvements and alternatives falling within the spirit and principles of the present disclosure.

Claims

1. A multi-factor plum rainfall prediction method based on self-adaptive graph structure and reinforcement integration is characterized by comprising the following steps:

Wherein,is a predicted value, Y _i Is the true tag value.

2. The prediction method according to claim 1, wherein the method for constructing the multi-factor adjacency matrix comprises the steps of:

constructing two groups of adaptively adjustable factor weight matrixesAnd->Where c represents the dimension of a parameter that can be learned in the weight matrix, the adjacency matrix is calculated using the following formula:

3. the prediction method according to claim 1, wherein the formula for normalizing the historical data of the factor related to plum rain is as follows:

4. a prediction method according to any one of claims 1-3, characterized in that the prediction model consists of four layers:

the first layer is input layer, and the input data is preprocessed data setAnd an adjacency matrix A;

the second layer is a graph roll-up neural network layer;

the third layer is a prediction network layer and comprises a gating circulation unit, a time convolution network and a Transformer neural network;

the fourth layer is a network integration layer, and three learnable weights w are set ₁ ，w ₂ ，w ₃ To weight sum the output of the prediction network layer to obtain the final prediction result, wherein w ₁ Is the weight of GRU, w ₂ Is the weight of TCN, w ₃ Is the weight of the transducer.

5. The prediction method of claim 4 wherein the graph rolling neural network layer is to collect data setsThe two-layer graph rolling operation is carried out with the adjacent matrix A, and the specific implementation mode is as follows:

6. The prediction method according to claim 4, wherein in the prediction network layer, the GRU and the fransformer are both 1 layer, and since the historical time step is 12, the prediction step is 1, the time convolution network is set to 5 layers, the convolution kernel sizes are all 3, the convolution kernel numbers are respectively 16, 32, 32, 40, 40, and the expansion coefficients are respectively 1,2,4,4,2, and adding a convolution layer with a size of 1×1 convolution kernel converts the output dimension of the time convolution network into 151×1×1.

7. The prediction method according to claim 4, wherein the learning weights of the network integration layer are trained by reinforcement learning, and the attenuation coefficient λ is set to 0.5, the learning rate γ is set to 0.95, and the maximum number of iterations is set to 50, specifically comprising:

constructing a state matrix S and an action matrix a, wherein the state matrix is the weight of three deep networks in a network integration layer, and the action matrix is the action of weight adjustment:

S＝[w ₁ ,w ₂ ,w ₃ ]#(4)

a＝[Δw ₁ ,Δw ₂ ,Δw ₃ ]#(5)

Constructing a loss function L, a reward function R and an evaluation function Q;

training the agent according to the training set, and executing an action matrix a by the agent according to the current state S;

calculating a loss function L to obtain rewards R, and making a next strategy;

wherein A (i) is the actual measurement data in the training set,predictive data in the training set;

calculating an evaluation function Q, and updating a Q table:

Q _m+1 (S _m ，a _m )＝Q _m (S _m ，a _m )+γ _m (R(S _m ，a _m )+λmaxQ _m (S _m+1 ，a _m+1 )-Q _m (S _m ，a _m ))#(8)

where λ is the decay coefficient and γ is the learning rate;

repeating the steps until the iteration stop condition is met, namely, the state matrix S is the optimal weight of the current three depth networks.

8. The prediction system of the prediction method according to any one of claims 1 to 7, characterized by comprising:

and a diagram construction module: the historical data for collecting factors related to plum rain is normalized, and the association relation among a plurality of factors is mined, so that a multi-factor adjacency matrix is constructed

wherein,is a predicted value, Y _i Is the true tag value.

9. A computer-readable storage medium, on which a computer program is stored, characterized in that the program, when executed, implements the steps of the method for predicting prune rainfall as claimed in any one of claims 1 to 7.

10. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the method for predicting the amount of prune rain as claimed in any one of claims 1-7 when the program is executed by the processor.