Research On Multi-scale Hydrological Drought Prediction In China Based On MLSTM Improved Deep Neural Network

Drought is a natural disaster that occurs in a wide area and causes serious losses.Drought is characterized by multiple temporal and spatial scales.Drought is a growing problem as the water cycle is gradually altered by climate change and human activities.The water cycle is gradually changing due to climate change and human activities,and the problem of drought is becoming more and more serious.China has a vast territory and complex terrain,at the same time,due to the influence of the monsoon circulation,China’s climate is complex,resulting in frequent disasters such as drought.With the influence of climate warming and human activities,the regional water cycle has intensified,resulting in an increase in the frequency,scope and intensity of droughts,posing severe challenges to the sustainable development of water resources and the high-quality construction of the ecological environment,which serious threat to China’s food security,water security,energy security and ecological health.Accurate drought prediction and early warning are important means for effective drought response and mitigation.However,the accuracy of current medium and long-term drought forecasts is low,and accurate drought forecasting is also a worldwide problem.Therefore,it is urgent to carry out the analysis of the temporal and spatial evolution of drought in China and the prediction of multi-temporal and spatial scales,in order to provide scientific theoretical support for early warning of drought and scientific drought prevention and disaster reduction in China.This paper takes China as the research object.Based on the runoff product with 0.25° precision,the multi-scale temporal and spatial evolution law of hydrological drought in China and the distribution of disaster risk levels were analyzed by using mathematical statistical methods such as standardized drought index and trend analysis.The predictability of hydrological droughts at multiple spatial and temporal scales in China was systematically explored.The main research results obtained in the thesis are as follows:(1)Using the SRI threshold method,the hydrological drought events in my country were identified,the spatio-temporal evolution law of hydrological drought in my country was revealed,and the distribution of hazard levels caused by hydrological drought was clarified.The research shows that:the multi-time scale hydrological drought in the Yangtze River Basin has no significant change trend,and the hydrological drought in the basin is generally light.The multi-time scale hydrological drought in the southeastern river basins,the Haihe River basin,the Songliao River basin and the Pearl River basin showed an overall aggravating trend,but the hydrological drought conditions in the basins were generally lighter.The hydrological drought situation on the 1-month and 3-month scales of the Huaihe River Basin showed an overall increasing trend,and the hydrological drought situation on the 6-month and 12-month scales did not change significantly,and the hydrological drought in the basin was more serious.The 1month hydrological drought in the Yellow River Basin and the Southwest River Basins generally showed a decreasing trend,and the hydrological drought in the 3-month,6-month and 12-month hydrological drought did not change significantly,and the hydrological drought in the basin was relatively serious.The hydrological drought situation in the inland river basins on multiple time scales showed a decreasing trend,but the drought situation in the inland river basin was the most severe.The disaster risk level of hydrological drought in China is generally characterized by a pattern of high in the west and low in the east,and high in the north and low in the south.Among them,the risk of hydrological drought in the northwest and northeast is generally higher,and the risk of hydrological drought in the south,southwest and southeast is higher.The disaster risk is generally low.The hazard risk level of SRI1 on the 1-month scale in most areas of China is high risk;the hazard risk level of SRI3 and SRI6 on the 3-month and 6-month scales in most areas is medium risk;the 12-month scale The hazard hazard level of SRI12 is medium risk(Songliao River Basin,Inland River Basin,Yellow River Basin,Haihe River Basin,Pearl River Basin,Southeast Basin and Yangtze River Basin)and low risk(Huaihe River Basin and Southwest River Basin).(2)Considering the characteristics of large volume and long time span of gridded hydrological data,and the long training period of deep neural network(LSTM)pixel by pixel.A novel data preprocessing strategy based on spatiotemporal random sampling was proposed,coupled with a deep neural network to construct an improved deep neural network hydrological drought prediction model(mLSTM).The results show that the mLSTM model reduces the time required to complete a national 0.25° precision grid hydrological drought prediction from 42.5 hours to 3 hours,and realizes the high-precision grid hydrological drought prediction in China.The model has a strong ability to predict the daily-scale hydrological drought in China on multiple time scales(SRI1,SRI3,SRI6 and SRI12).The model also has a certain ability to predict the multi-time-scale monthly hydrological drought in China.At the same time,the prediction effect of the model on the long-term prediction of daily-scale hydrological drought in China is ideal,which proves the rationality of mLSTM,and the model has strong robustness and generalization ability.(3)Using the mLSTM model to predict multi-time-scale daily hydrological drought in China,the predictability of daily-scale hydrological drought in my country has been proved,and its temporal and spatial distribution and dynamic evolution law have been revealed.The mLSTM model has a high accuracy in predicting daily-scale hydrological drought in China.Except for the southeastern river basins,there is a certain correlation between hydrological drought and atmospheric circulation factors in China and other eight river basins.After inputting the integrated atmospheric circulation factor,the prediction accuracy and hit rate of the daily-scale hydrological drought in China have been improved.The predictability of the atmospheric circulation factor provides a certain degree of predictability for the daily-scale hydrological drought in China,indicating that the overall daily-scale hydrological drought in China is predictable(the increase of mean of NSE values and rhit values is 22.92%and 19.78%).The Songliao River Basin has the highest predictability(39.51%and 20.71%),followed by the Southwest River Basin(35.31%and 27.15%)and the Inland River Basin(34.32%and 18.90%).The Yangtze River Basin(8.22%and 16.79%)has the weakest predictability.The predictability of other river basins is in the middle,the Yellow River Basin(27.52%and 13.68%),Huaihe River Basin(21.27%and 17.41%),Haihe River Basin(20.46%and 18.16%),Pearl River Basin(15.11%and 20.71%)and Southeast Asia.watershed(12.68%and 17.39%).(4)Using the mLSTM model to predict the multi-temporal and spatial monthly hydrological drought in my country,the predictability of the monthly hydrological drought in my country has been proved,and its temporal and spatial distribution and dynamic evolution have been revealed.The results show that the prediction scheme using the national hydrological drought index time series as input can improve the prediction accuracy of monthly hydrological drought in China.The prediction accuracy of the mLSTM model for monthly hydrological drought increases with the increase of time scale.After adding the atmospheric circulation factor to the model input,the prediction accuracy and hit rate of monthly hydrological drought are improved,and the false alarm rate is reduced.The average NSE values in China and the nine major river basins have increased by more than 100%.It is easy to see that the monthly-scale hydrological drought in my country is predictable in general(the average increase in rhit is 70.29%).The southwestern river basins have the strongest predictability(an increase of 106.07%).The Pearl River Basin and the Yangtze River Basin were second(increased by 88.00%).And the southeastern river basins have the weakest predictability(an increase of 23.37%).The predictability of other river basins is in the order of Songliao River Basin(67.60%),Haihe River Basin(63.15%),Yellow River Basin(63.05%),Huai River Basin(59.51%)and Inland River Basin(49.63%).(5)Using the mLSTM model to study the long-term forecasting period of daily-scale hydrological drought in my country in three periods,revealing the predictability of the long-term forecast period of daily-scale hydrological drought in China.The results show that the performance of mLSTM for the long-period forecast of daily-scale hydrological drought in my country is relatively ideal,and it can be competent for the 15-day forecast period of daily-scale hydrological drought in China.The prediction accuracy grades of the model for the longforecast period prediction of daily-scale hydrological drought in my country in the three periods are all above grade C.The prediction effect of the mLSTM model on the long-foreseeing period of daily-scale hydrological drought in China in the three periods gradually decreases with the increase of the forecast period(the average decline rate of NSE in 1948-1970 was 2.38%/day,and it was 2.27%/day in 1971-1992,1993-2014 was 2.58%/day).After inputting the fused atmospheric circulation factor,the prediction accuracy of the corresponding forecast period was improved,and the decay rate of the prediction accuracy was slowed down(the average decline rates of NSE in the three periods were all reduced by 2.10%,2.64%and 13.57%respectively).The predictability of the long forecast period of daily hydrological drought in China showed an overall trend of first increasing and then decreasing.The predictability of the long forecast period of daily hydrological drought was the strongest between 1971 and 1992(the average decline rate of NSE was 2.21%/day),and in 1993-The period 2014 was the next most predictable(2.23%/day);the period 1948-1970 was the most difficult and least predictable(2.33%/day).(6)The dominant factors in the spatial differentiation of hydrological drought predictability in China were detected using geodetectors.The results show that two meteorological factors,temperature and rainfall,are the dominant factors in the spatial differentiation of hydrological drought predictability in my country and nine major river basins.The dominant effect of the two-factor interaction on the spatial differentiation of hydrological drought measurability is stronger than that of a single factor.The interaction between temperature and precipitation is of the spatial differentiation of hydrological drought predictability in China and the southeastern river basins,the Yellow River basin,the Songliao river basin and the southwestern river basins.The interaction of temperature,precipitation and ndvi was the dominant factor for the spatial differentiation of hydrological drought predictability in the Yangtze River Basin,Hai River Basin,Inland River Basin and Pearl River Basin.The interaction of precipitation and elevation was the dominant factor for the spatial differentiation of predictability of hydrological drought in the Huaihe River Basin.