Objective forecast of maximum temperature is an important part in numerical weather prediction(NWP). The forecast uncertainty of near-surface meteorological elements is greater than that of upper atmospheric elements due to the impact of uncertainty in numerical forecasting models for sub-grid and boundary layer schemes.In recent years, meteorological observations expand rapidly, making traditional error correct method difficult to deal with the massive data. As a result, artificial intelligence has an increasingly obvious advantage in processing big data. Based on the fully connected neural network, four sensitivity experiments are designed in order to investigate the importance of auxiliary variable, time-lagged variable and the effectiveness of embedding layer in the neural network. The output products of high resolution(HRES) model of European Centre for Medium-Range Weather Forecasts(ECMWF) and the observations of basic meteorological elements of totally 2238 basic weather stations from 15 January 2015 to 31 December 2020 are employed. The training period is from 15 January 2015 to 31 December 2019, and the rest part is test period.The results show that the forecast error of daily maximum air temperature from the HRES in test period is reduced greatly by the sensitivity experiments, which add auxiliary variables, daily maximum air temperature with 1-2 lag days and embedding layer structures and their combination. The root mean square error is reduced by 29.72%-47.82% and the accuracy of temperature forecast are increased by 16.67%-38.89%, and the effects for Qinghai-Tibet Plateau is especially remarkable where the forecast error of HRES model is very high. It is preliminarily proved that the fully connected neural network with embedding layer has better overall performance than the raw fully connected neural network, and the features also affect the forecast errors and forecast skills of the model. Besides, the prediction error of neural network model with embedding layer is more stable when auxiliary variables and lag time variables are added. Positive forecasting techniques are available for almost all stations in the study, and it is possible to reduce the mean absolute error to less than 1℃ at many stations.

Rapid-refresh Multi-Scale Analysis and Prediction System-Integration (RMAPS_IN) is an important tool for Beijing, Hebei and other meteorological departments to make rapid-updated and refined precipitation nowcasting. The precipitation analysis products of the system are based on automatic station observation and radar quantitative precipitation estimation data, while 0-2 h forecast products are obtained by extrapolation based on the analysis products. To study the applicability of different extrapolation methods in RMAPS_IN, the precipitation events of different weather systems from 2019 to 2020 are analyzed, using cross correlation method and optical flow method to conduct a 0-2 h extrapolation nowcasting test based on the RMAPS_IN precipitation analysis products. The cross correlation method uses classic optimal correlation coefficient calculation scheme, while the optical flow method employs the Farneback dense optical flow calculation scheme in the OpenCV function library. According to the characteristics of the regional weather systems, the precipitation events are divided into five types: Low trough cold front precipitation, low vortex precipitation, typhoon precipitation, cyclone precipitation, and warm shear line precipitation. The sample size of each precipitation type is 2108, 1448, 1058, 260, and 140, respectively. The batch test results show that the extrapolated vectors by the cross correlation method and optical flow method have a certain difference in magnitude and direction. The direct difference has a clear correspondence with the position of the weather system that affects precipitation, and is more obviously affected by the geographical location. For typhoon precipitation, the difference in direction is distributed in an arc band, while for other 4 types of precipitation, the difference is large in the northwest and small in the southeast. In terms of forecasting effect, the cross correlation method is generally better than the optical flow method, especially when the forecast time exceeds 30 minutes, and the longer the lead time is, the more obvious the advantage is. But when the forecast time is 10 min, the optical flow method is better in the false alarm rate of low vortex precipitation, typhoon precipitation and warm shear line precipitation. In addition, the nowcasting method based on extrapolation has the best prediction effects on typhoon precipitation in Beijing-Tianjin-Hebei region, followed by warm shear line precipitation, low vortex precipitation, low trough cold front precipitation, and cyclone precipitation. It should be noted that in Beijing-Tianjin-Hebei region, cyclone precipitation and warm shear line precipitation rarely occurred in recent years, and the sample size of these two types of precipitation is significantly smaller than that of other types, so the relevant results are less representative.

To address the practical needs of objectively describing the uncertainty of rainfall-based landslides and the existing problems of single warning indicators and subjective forecasting methods in the meteorological disaster early warning business, landslide disaster data from 2014 to 2020 and multi-source used precipitation analysis data are investigated to construct a regional rainfall-induced landslides probability forecasting model. Machine learning classification algorithms is implemented through key steps such as sample construction, model training, parameter optimization and forecast output to explore the feasibility of different types of algorithms in identifying landslides-causing rainfall processes. A training sample set construction method based on the positive samples, the negative samples are obtained by sampling under spatial-temporal limitation. The evaluation of different machine learning classification algorithms using the sample set shows that linear discriminant analysis algorithm has the highest accuracy(0.863) and the best generalization ability(area under the receiver operating characteristic curve is 0.886) without over-fitting problem, followed by the logistic regression algorithm and the K-nearest neighbor algorithm. In the probabilistic forecasting test for the cases of rainfall-induced landslides in 2021, all of three algorithms can extract and learn the conditional features and have certain ability to identify the rainfall processes which induce landslides. K-nearest neighbor algorithms and logistic regression algorithms have a relatively large range of probabilistic forecasting high value areas, which are prone to false alarm results. The probability forecast of the linear discriminant analysis algorithms is more convergent in the range of the high value area, and it can extract local rainfall information better, but it outputs unnecessary low-value probability forecasts in non-rainfall central area. The rainfall-induced landslides probability prediction model based on the machine learning classification algorithm comprehensively considers the coupling effect of the underlying surface factor and the rainfall factor, which is better than the commonly used critical threshold model that assumes the occurrence of landslides in the forecast area is only related to rainfall. The application results show that the machine learning classification algorithm model makes up for the shortcomings of existing forecasting models that are less likely to reflect the influence of the surface environment, so it is an important way to improve the performance of landslides forecasting and warning.

Precipitation can be induced by various weather systems and a series of complex physical processes, so its prediction is relatively difficult in weather forecasting. Due to the limitation of numerical model, the prediction error is inevitable. It is a hot topic in meteorological research and operation to explore a more effective method to correct the model product, and to improve the interpretation and applicability. To explore a more effective model product error correction method, a combination of correction methods is put forward, based on singular value decomposition (SVD) and machine learning, including multiple linear regression, LASSO regression and Ridge regression. The results are compared with the traditional matrix coefficient method, and then correction models are tested in pre-flood season precipitation forecast in South China, by correcting European Centre for Medium-Range Weather Forecasts (EC) product. The result shows that the proposed correction models combining SVD and machine learning can effectively reduce the error of EC product. The maximum optimization rate root mean square error is 4.2%, and more than 69% of the stations are optimized to different degrees. These correction models have better robustness to deal with the problem of collinearity between factors, and the correction effect is better than that of the traditional matrix coefficient method. Furthermore, the weighted integration of multiple correction models is carried out by assigning different weights to different models, and the root mean square error by the integrated approach in South China is smaller than EC product and any single correction model. It shows that the weighted ensemble method can better integrate the advantages of multiple correction models and enlarge the advantages. For the weighted ensemble of multiple correction models, it is not only better than the precipitation prediction results of EC product, but also better than any one of the integrated correction models. Its optimization rate of root mean squared error can achieve 5.7%, and more than 77% of the stations are optimized to different degrees.

Two strong supercell tornados hit Shangzhi Acheng of Harbin and Meilisi of Qiqihar, Heilongjiang Province on 1 June ("6·1" Tornado) and 9 June ("6·9" Tornado) in 2021. Using the conventional meteorological observations and Doppler weather radar data, the multiscale characteristics of two events are analyzed.Both events occur in the southeast quadrant of northeastern cold vortex. The left outlet of the upper-level jet stream and the southerly jet stream at lower-level are conducive to the development of vertical movement and the transport of warm-wet air. The temperature difference between 850 hPa and 500 hPa exceeds 30℃. Two storms are both triggered by mesoscale dry-lines and convergence lines. The pseudo-cold fronts, which generate from the mesoscale warm front and the cold pool coming from the thunderstorm outflow, are beneficial to the development and maintenance of tornados. Tornados appear on the wet part of the junctions between the pseudo-cold front and dry line, and in the front of cold pool. The parent storms of tornados rapidly develop into supercells as they pass over water bodies such as reservoirs and wetlands. The warm-wet inflow gap indicates the development of hook echo. Medium to strong mesocyclones firstly appear at about 3 km high, and then go upwards and downwards, touchdown 5-10 minutes later. The tornados occur when hook echoes and mesocyclones appear simultaneously.There are also some differences between them. Short-time heavy rainfall occurs on 1 June and thunder-gust occurs on 9 June with typical sounding layer structures, but "6·1" Tornado is stronger. The atmospheric instabilities are dominated by cold advection at upper-level for "6·1" Tornado but warm advection at lower-level for "6·9" Tornado. Water vapor and vertical velocity of "6·1" Tornado is more beneficial to the development of supercell than those of "6·9" Tornado. For the vertical wind shears of 0-1 km and 0-6 km, the corrected lifting condensation level and convective available potential energy (CAPE) are 12 m·s^-1, 18 m·s^-1, 770 m and 420 J·kg^-1 for "6·1" Tornado, and 10 m·s^-1, 33 m·s^-1, 1100 m and 2500 J·kg^-1 for "6·9" Tornado. The stronger 0-1 km wind shears and the lower corrected lifting condensation level show the possibility of intense tornado. CAPE may be underestimated because of the spatiotemporal resolution limitation for soundings.The main cause for the long duration of "6·1" Tornado is that the mesoscale vortex at 3 km altitude maintains due to the continuous warm-wet inflow. However, the strong mesocyclone of "6·9" Tornado doesn't last that long.