A comparative study of extensive machine learning models for predicting long-term monthly rainfall with an ensemble of climatic and meteorological predictors

Rainfall prediction is of vital importance in water resources management. Accurate long-term rainfall prediction remains an open and challenging problem. Machine learning techniques, as an increasingly popular approach, provide an attractive alternative to traditional methods. The main objective of this study was to improve the prediction accuracy of machine learning-based methods for monthly rainfall, and to improve the understanding of the role of large-scale climatic variables and local meteorological variables in rainfall prediction. One regression model autoregressive integrated moving average model (ARIMA) and five state-of-the-art machine learning algorithms, including artificial neural networks, support vector machine, random forest (RF), gradient boosting regression, and dual-stage attention-based recurrent neural network, were implemented for monthly rainfall prediction over 25 stations in the East China region. The results showed that the ML models outperformed ARIMA model, and RF relatively outperformed other models. Local meteorological variables, humidity, and sunshine duration, were the most important predictors in improving prediction accuracy. 4-month lagged Western North Pacific Monsoon had higher importance than other large-scale climatic variables. The overall output of rainfall prediction was scalable and could be readily generalized to other regions.     

Genetic algorithms for the classification and prediction of precipitation occurrence

Abstract

Using an approach similar to the biological processes of natural selection and evolution, the genetic algorithm (GA) is a nonconventional optimum search technique. Genetic algorithms have the ability to search large and complex decision spaces and handle nonconvexities. In this paper, the GA is applied for solving the optimum classification of rainy and non-rainy day occurrences based on vertical velocity, dewpoint depression, temperature and humidity data. The problem involves finding optimum classification based on known data, training the future prediction system and then making reliable predictions for rainfall occurrences which have significance in agricultural, transportation, water resources and tourism activities. Various statistical approaches require restrictive assumptions such as stationarity, homogeneity and normal probability distribution of the hydrological variables concerned. The GAs do not require any of these assumptions in their applications. The GA approach for the occurrence classifications and predictions is presented in steps and then the application of the methodology is shown for precipitation occurrence (non-occurrence) data. It has been shown that GAs give better results than classical approaches such as discriminant analysis. The application of the methodology is presented independently for the precipitation event occurrences and forecasting at the Lake Van station in eastern Turkey. Finally, the amounts of precipitation are predicted with a model similar to a third order Markov model whose parameters are estimated by the GA technique.     

Application of newly developed ensemble machine learning models for daily suspended sediment load prediction and related uncertainty analysis

ABSTRACT

Ensemble machine learning models have been widely used in hydro-systems modeling as robust prediction tools that combine multiple decision trees. In this study, three newly developed ensemble machine learning models, namely gradient boost regression (GBR), AdaBoost regression (ABR) and random forest regression (RFR) are proposed for prediction of suspended sediment load (SSL), and their prediction performance and related uncertainty are assessed. The SSL of the Mississippi River, which is one of the major world rivers and is significantly affected by sedimentation, is predicted based on daily values of river discharge (Q) and suspended sediment concentration (SSC). Based on performance metrics and visualization, the RFR model shows a slight lead in prediction performance. The uncertainty analysis also indicates that the input variable combination has more impact on the obtained predictions than the model structure selection.     

Incorporating synoptic-scale climate signals for streamflow modelling over the Mediterranean region using machine learning models

ABSTRACT

Understanding streamflow patterns by incorporating climate signal information can contribute remarkably to the knowledge of future local environmental flows. Three machine learning models, the multivariate adaptive regression splines (MARS), the M5 Model Tree and the least squares support vector machine (LSSVM) are established to predict the streamflow pattern over the Mediterranean region of Turkey (Besiri and Baykan stations). The structure of the predictive models is built using synoptic-scale climate signal information and river flow data from antecedent records. The predictive models are evaluated and assessed using quantitative and graphical statistics. The correlation analysis demonstrates that the North Pacific (NP) and the East Central Tropical Pacific Sea Surface Temperature (Niño3.4) indices have a substantial influence on the streamflow patterns, in addition to the historical information obtained from the river flow data. The model results reveal the utility of the LSSVM model over the other models through incorporating climate signal information for modelling streamflow.     

地震动强度参数估计的可解释性与不确定度机器学习模型

准确预测地震动强度参数(峰值加速度PGA、峰值速度PGV等)对于震后应急和地震危险性概率分析至关重要.作为地震动强度参数预测的新手段, 机器学习算法具有优势, 但也存在可解释性差和难给出预测结果不确定度的问题.本文提出采用自然梯度提升(NGBoost)算法在预测结果的同时提供其不确定度, 并结合SHAP值解释机器学习模型.基于NGA-WEST2强震动数据库, 本文训练出了适合预测活跃构造区地壳地震的PGA和PGV概率密度分布的机器学习模型.测试集数据PGA和PGV的预测值与真实值的相关系数可达0.972和0.984, 并可给出预测结果的合理概率密度分布.通过SHAP值, 我们从数据角度弄清了各输入特征(矩震级M_W、Joyner-Boore断层距R_jb、地下30 m平均S波速度V_S30、滑动角Rake、断层倾角Dip、断层顶部深度Z_TOR和V_S达到2.5 km·s^-1时的深度Z_2.5)对机器学习模型预测结果的影响机理.SHAP值显示, 基于NGBoost算法的机器学习模型的预测方式基本与物理原理相符, 说明了机器学习模型的合理性.SHAP值还揭示出一些以往研究忽视的现象: (1)对于活跃构造区地壳地震, 破裂深度较浅(Z_TOR＜~5 km)时, Z_TOR的SHAP值低于破裂深度较深(Z_TOR＞~5 km)时的值, 表明浅部破裂可能主要受速度强化控制, 地震动强度较弱.并且Z_TOR的SHAP值随Z_TOR值增大而减小, 表明地震动强度可能还受破裂深度变化引起的几何衰减变化影响; (2)破裂深度较深时, Z_TOR的SHAP值随Z_TOR值增大而增大, 表明深部破裂的地震动强度可能受和破裂深度变化相关的应力降或品质因子Q的变化影响; (3)Z_2.5较小(Z_2.5＜~1 km)时, Z_2.5的SHAP值的变化规律对于PGA和PGV预测是相反的, 表明加速度和速度频率不同, 受浅层沉积物厚度变化引起的共振频率变化影响不同.

     

动力和统计预测信息融合预测方法及对我国夏季降水预测的检验

短期气候预测中如何将气候模式和统计方法的预测结果科学、客观的集成起来,一直是非常重要的问题.本文针对动力模式和统计方法预测结果相结合的问题,引入资料同化中信息融合的思想,采用最优内插同化方法,实现了动力模式和统计季节降水预测结果的融合.检验表明,对1982-2015年我国夏季降水百分率的回报,融合预测结果与观测的平均空间相关系数可达0.44,分别较统计预测和CFSv2模式统计降尺度订正的技巧提高了0.1左右,而均方根误差较两者可以降低5%~20%.可见,该方法可以进一步提升对我国夏季降水的预测技巧,具有显著的业务应用价值.

     

基于迁移学习的地球物理测井储层参数预测方法研究

随着大数据和机器学习的成熟和推广应用,人工神经网络在地球物理测井预测储层参数中得到重视.本文引入迁移学习进行测井储层参数预测,以孔隙度预测神经网络模型和孔隙度含水饱和度联合预测神经网络模型为基础模型,分别以渗透率及含水饱和度预测作为目标任务进行迁移学习,以提升储层参数预测效果和效率.文中详细阐述了基于迁移学习的测井储层参数预测方法,并使用64口井的测井数据进行储层参数预测效果分析.结果表明,使用迁移学习后,渗透率模型预测效果最高可以提升58.3%;含水饱和度模型预测效果最高可以提升近40%,且最大可以节省60%的计算资源;以孔隙度预测模型为基础模型时更适合使用参数冻结的训练方式,以孔隙度含水饱和度联合预测模型为基础模型时更适合使用参数微调的训练方式.

     

机器学习算法在发震构造形态特征及地震迁移规律研究中的应用前景

对发震构造形态特征和地震迁移规律的研究,有利于震情形势研判和震源物理的研究,也可为后期三维地壳介质结构建模提供数据支持.本研究结合大量中小地震的分布特征规律,利用机器学习的相关算法,提取断层形态特征和地震迁移规律.     

Multi-step prediction of strong earthquake ground motions and seismic responses of SDOF systems based on EMD-ELM method

This paper proposes a new multi-step prediction method of EMD-ELM (empirical mode decomposition-extreme learning machine) to achieve the short-term prediction of strong earthquake ground motions. Firstly, the acceleration time histories of near-fault ground motions with nonstationary property are decomposed into several components of intrinsic mode functions (IMFs) with different characteristic scales by the technique of EMD. Subsequently, the ELM method is utilized to predict the IMF components. Moreover, the predicted values of each IMF component are superimposed, and the short-term prediction of ground motions is attained with low error. The predicted results of near-fault acceleration records demonstrate that the EMD-ELM method can realize multi-step prediction of acceleration records with relatively high accuracy. Finally, the elastic and inelastic acceleration, velocity and displacement responses of single degree of freedom (SDOF) systems are also predicted with satisfactory accuracy by EMD-ELM method.     

多因素影响的建筑物群体震害预测方法研究

为了简化震害预测工作,提出了一种多因素影响的建筑物群体震害预测方法。首先,将已有数据库中的资料按不同相似度进行分类,从中选取所需要的样本数据。然后将所选取的样本数据按不同影响因素分类,分别求出考虑各影响因子下的震害矩阵,再由房屋普查资料得出预测区考虑各影响因素时各影响因子下的房屋的建筑面积,并将建筑面积比例作为各影响因子的权重,最终得出预测区某种结构类型整体的震害矩阵。利用文中方法建立了厦门市多层砌体结构的震害矩阵,与厦门市采用单体抽样法得出的震害矩阵相比较,其平均震害指数最大差值不大于0.041,验证了此方法的可行性。     

基于半监督学习的井震联合储层横向孔隙度预测方法

传统地震储层预测技术一般基于弹性参数反演和岩石物理建模的级联流程实现储层孔隙度预测, 其预测精度受到波动理论和岩石物理理论的近似假设、初始模型和二次反演累积误差等因素的影响.为缓解这些问题, 本文提出了一种基于双向门控递归单元神经网络的半监督学习井震联合孔隙度预测方法, 实现从地震数据直接预测储层横向孔隙度.通过少量的地震测井样本标签对和多目标函数约束建立智能化多尺度多信息融合孔隙度预测模型, 实现地震数据到孔隙度, 孔隙度再到生成地震数据的闭环映射.此外, 在网络模型每次迭代更新的过程中随机引入非井旁地震道参与网络训练, 非井旁地震道的波形匹配能在一定程度上保证井间孔隙度的预测精度.模型数据和实际数据测试结果表明, 本文提出的方法相比于有监督学习孔隙度预测方法能进一步提高储层孔隙度的预测准确性和横向连续性, 获得较为可靠的储层物性参数的空间分布.

     

水氡浓度和环境参数的分层神经网络研究

采用分层神经网络（LNN）分析地下水的氡浓度,试图给出氡浓度和环境参数之间的函数关系。由于环境（例如：降雨量）对水氡浓度的影响可能是非线性的,与目前时间脉冲响应线性计算方法相比,该方法能够较准确的估计环境参数造成的氡浓度变化。     

Deep learning approaches for improving prediction of daily stream temperature in data-scarce,unmonitored, and dammed basins

Basin-centric long short-term memory (LSTM) network models have recently been shown to be an exceptionally powerful tool for stream temperature (T_s) temporal prediction (training in one period and predicting in another period at the same sites). However, spatial extrapolation is a well-known challenge to modelling T_s and it is uncertain how an LSTM-based daily T_s model will perform in unmonitored or dammed basins. Here we compiled a new benchmark dataset consisting of >400 basins across the contiguous United States in different data availability groups (DAG, meaning the daily sampling frequency) with and without major dams, and studied how to assemble suitable training datasets for predictions in basins with or without temperature monitoring. For prediction in unmonitored basins (PUB), LSTM produced a root-mean-square error (RMSE) of 1.129°C and an R² of 0.983. While these metrics declined from LSTM's temporal prediction performance, they far surpassed traditional models' PUB values, and were competitive with traditional models' temporal prediction on calibrated sites. Even for unmonitored basins with major reservoirs, we obtained a median RMSE of 1.202°C and an R² of 0.984. For temporal prediction, the most suitable training set was the matching DAG that the basin could be grouped into (for example, the 60% DAG was most suitable for a basin with 61% data availability). However, for PUB, a training dataset including all basins with data was consistently preferred. An input-selection ensemble moderately mitigated attribute overfitting. Our results indicate there are influential latent processes not sufficiently described by the inputs (e.g., geology, wetland covers), but temporal fluctuations can still be predicted well, and LSTM appears to be a highly accurate T_s modelling tool even for spatial extrapolation.     

Ensemble rainfall forecasting with numerical weather prediction and radar‐based nowcasting models

The overall objective of this study is to improve the forecasting accuracy of the precipitation in the Singapore region by means of both rainfall forecasting and nowcasting. Numerical Weather Predication (NWP) and radar‐based rainfall nowcasting are two important sources for quantitative precipitation forecast. In this paper, an attempt to combine rainfall prediction from a high‐resolution mesoscale weather model and a radar‐based rainfall model was performed. Two rainfall forecasting methods were selected and examined: (i) the weather research and forecasting model (WRF); and (ii) a translation model (TM). The WRF model, at a high spatial resolution, was run over the domain of interest using the Global Forecast System data as initializing fields. Some heavy rainfall events were selected from data record and used to test the forecast capability of WRF and TM. Results obtained from TM and WRF were then combined together to form an ensemble rainfall forecasting model, by assigning weights of 0.7 and 0.3 weights to TM and WRF, respectively. This paper presented results from WRF and TM, and the resulting ensemble rainfall forecasting; comparisons with station data were conducted as well. It was shown that results from WRF are very useful as advisory of anticipated heavy rainfall events, whereas those from TM, which used information of rain cells already appearing on the radar screen, were more accurate for rainfall nowcasting as expected. The ensemble rainfall forecasting compares reasonably well with the station observation data. Copyright © 2012 John Wiley & Sons, Ltd.     

基于岩石物理转换模型的苏里格致密砂岩储层测井评价与产能预测

苏里格地区致密砂岩储层勘探开发难度很大,测井解释遇到了储层参数计算和产能预测不准确的问题.本文针对这些难点,基于已有的转换模型,推导并验证了T₂-I和T₂-K_r转换模型,提出了采用转换模型应用于测井综合解释和产能预测的方法.基于实验数据设定孔径大于1 μm的孔隙为大孔,中孔孔径在1 μm和0.03 μm之间,小孔孔径为0.03 μm以下.并建立了不同孔径范围模型中关键参数α与核磁参数T_2lm的关系,这为转换模型在测井解释中的应用提供了必要条件.本文在苏里格西区分别应用T₂-I和T₂-K_r转换模型求取含水饱和度和相对渗透率曲线,并进行产能预测,处理结果显示该方法具有很好的应用效果.

     

一种新颖的探地雷达快速正演模拟及埋地目标探测机器学习方法

探地雷达正演模拟在真实雷达数据解译及全波形反演中扮演着重要的角色,针对传统探地雷达（Ground Penetrating Radar,GPR）正演模拟计算量巨大、耗时、不利于实时探测等问题,提出一种基于机器学习框架的近实时GPR正演模拟方法.以混凝土中的钢筋探测作为GPR应用场景,混凝土的含水量、钢筋半径及埋地深度作为模型参数,利用时域有限差分数值模拟散射回波信号;运用主成分分析对回波数据进行降维处理得到相应的主成分权值系数,并以此作为机器学习网络的输出;设计了一种基于随机森林的多层循环网络架构和学习策略,不仅充分挖掘学习模型参数和主成分权值系数之间的内在因果关系,也共享主成分间的相互联系,并具有对每个预测主成分完善和修正的功能,以此实现基于机器学习的探地雷达快速正演模拟,与传统机器学习相比,有效提高了正演模拟的精度.在此基础上将两个深度神经网络与随机森林相结合,以回波数据主成分系数为输入,建立了基于机器学习的场景参数预测模型,实现了近实时的埋地目标探测,预测的混凝土含水量最大误差为2%,钢筋埋地深度最大误差为6.7%.

     

基于卫星数据和NRLMSISE-00模型的低轨道大气密度预报修正方法

NRLMSISE-00大气模型广泛应用于航天器定轨和预测等方面,但存在着较大的误差,尤其是在短期变化方面.为了提高低轨道大气密度短期预报的精度,我们提出了一种基于实测数据对NRLMSISE-00大气模型密度结果进行修正预报的方法:利用GRACE(Gravity Recovery and Climate Experiment)和CHAMP(Challenging Mini-Satellite Payload)卫星2002—2008年大气密度探测数据对NRLMSISE-00模型进行误差分析,获得模型的修正因子,再对模型的大气密度结果进行修正.采用该修正方法对GRACE-A和CHAMP卫星轨道上的大气密度进行3天短期预报试验验证,结果表明可显著提高大气密度的预报精度,在太阳活动低年,修正后的大气密度预报误差比NRLMSISE-00模型误差降低50%以上.     

基于机器学习的水-气界面CO2、CH4扩散通量预测及影响因素分析——以三峡水库为例

传统的水-气界面温室气体通量的监测方法具有诸多局限,对其影响因素的分析也大多基于数学统计层面。对此,本研究提供了一种较为新颖的研究和分析方法——基于机器学习的数据预测和分析。本研究采用2种经典机器学习算法——随机森林(RF)和支持向量机(SVM)和2种深度学习算法——卷积神经网络(CNN)和长短时记忆神经网络(LSTM),通过环境因素预测水库水-气界面CO₂和CH₄扩散通量。此外,采用RF中的特征重要性评估和经典算法决策树(DT),对环境因素和水库温室气体扩散通量的关系进行了全新角度的数据挖掘和分析。结果表明:深度学习算法的预测效果均较好,经典机器学习算法中RF预测效果显著优于SVM。LSTM和RF分别产生了最优的CO₂扩散通量和CH₄扩散通量的预测精度,均方根误差(RMSE)分别为0.424 mmol/(m²·h)和0.140μmol/(m²·h),预测值与实测值的R²分别为0.960和0.758。RF的特征重要性评估表明沉积物因子...     

基于机器学习和台阵相关性的水力压裂微地震事件自动识别及到时拾取

利用密集台阵对水力压裂微地震进行监测将有助于优化储层压裂、揭示断层活化.为满足密集台阵海量采集数据的处理需求, 本文建立了一种综合运用多种机器学习方法和台阵相关性的、无需人工干预的自动处理流程, 从而能够快速得到高质量的密集台阵震相到时目录.该综合策略包括: (1)利用迁移学习在连续波形中快速检测地震事件; (2)利用U型神经网络PhaseNet自动拾取P波、S波震相; (3)利用三重线性剔除法, 结合密集台阵到时相关性剔除异常到时数据和地震事件; (4)利用K-means和SVM两类机器学习算法, 进一步区分发震时刻接近的多个地震事件, 减小事件漏拾率.通过将该流程应用于四川盆地长宁—昭通页岩气开发区微地震监测数据, 并将自动处理结果与人工拾取结果进行比对发现, 二者在震级测定、定位以及走时成像结果等方面具有很好的一致性, 表明本文处理流程结果精度可达到手动处理精度.本文结果为密集台阵地震监测数据的高效、高精度处理提供了新思路.