Ensembles, multi-models
See:
# Resources
- Improving predictions of Ensembles and averaging. Reducing uncertainty on future predictions
- No one model predicts best all the time, for all variables
- Best predictor: Average predictions over all models
- Average prediction weights all models equally
- Weighted average prediction gives varying weights to each models based on past performances
- Adaptive weighted average prediction identifies current best predicting model vs one that quickly switching to other models
- Online Learning: Non stationary data. Learns the switching rate: level of non-stationarity
- Multi-model assessment of seasonal T and precipitation forecasts over Europe
- Best predictor: Average predictions over all models
- Ensemble Verification Metrics
- IPCC Expert Meeting on Assessing and Combining Multi Model Climate Projections â IPCC
# Talks
- #TALK Statistical post-processing of ensemble weather forecasts: Current developments and future directions (Tilmann Gneiting)
- #TALK Multi-model ensemble predictions on seasonal timescale
# Courses
# References
- #PAPER
Decomposition of the Continuous Ranked Probability Score for Ensemble Prediction Systems (Hersbach 2000)
- The CRPS can be seen as a ranked probability score with an infinite number of classes, each of zero width
- Alternatively, it can be interpreted as the integral of the Brier score over all possible threshold values for the parameter under consideration
- For a deterministic forecast system the CRPS reduces to the mean absolute error
- CRPS can be decomposed into three parts:
- reliability, is closely related to the rank histogram. The reliability should be zero for an ensemble system with the correct statistical properties.
- uncertainty, is the best achievable value of the continuous ranked probability score, in case only climatological information is available.
- resolution, expresses the superiority of a forecast system with respect to a forecast system based on climatology.
- #PAPER A multiple model assessment of seasonal climate forecast skill for applications (Lavers 2009)
- #PAPER Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles (Lakshminarayanan 2017)
- #PAPER Predicting Weather Forecast Uncertainty with Machine Learning (Scher 2018)
- #PAPER
Using multiâmodel ensembles of CMIP5 global climate models to reproduce observed monthly rainfall and temperature with machine learning methods in Australia (Wang 2018)
- https://agrivy.oss-cn-zhangjiakou.aliyuncs.com/papers_agrivy/webfiles/papers/2018-IJC-WANG-BIN.pdf
- The purpose of this study is to compare the capacity of four different multi-model ensemble (MME) methods (random forest, support vector machine, Bayesian model averaging and the arithmetic ensemble mean) in reproducing observed monthly rainfall and temperature
- #PAPER
Predicting Weather Uncertainty with Deep CNNs (Gronquist 2019)
- Ensembles uncertainty estimation with DL
- WF uncertainty quantification using ensembles prediction system (nonparametrics stats on multiple perturbed simulations)
- Intensive simulations (dozens and up to 50)
- Each ensemble member is perturbed, the STDDEV of the embers can be used to identify the uncertainty of a HRes forecast
- Data: ERA5 (similar to production NWP data) Reanalysis by ECMWF with weather data reanalysis from 1979.
- Ensemble of 9 perturbed trajectories and a single unperturbed (control) one. Mapped to 0.5 degree resolution, 37 pressure levels. Temperature prediction.
- Subset of IPs that have an influence on Temp: zonal and meridional wind, geopotential, temperature, relative humidity and the factions of cloud cover
- Cropping for EU and Atlantic, 40 lat by 136 lon
- 7 pressure levels including 500 hPa, 850 hPa
- Temporally: 2000 â 2011, between 0600 and 1800 UTC with forecasts made for 3h and 6h into the future
- Standardize the data for each pressure level and parameter
- Second dataset: ENS10, re-forecast with 10 perturbed members, 24 h intervals. 2 times per week the last 20 years. Similar to the operational 51-member ensemble, but coarser resolution of 0.5 degrees
- Model:Â Weight sharing on each pressure level separately (“full”), more representational power but more parameters
- Point-wise affine transforms per pressure level after each convolution (“affine”) 2D convs followed by 1D vertical convs (“separable”)
- Temporal trends (can the temporal progression of the spread be learned?): data of the spread of all 10 trajectories at times 0h, 3h and 6h Tried CLSTMs At the end, treating time sequences as additional channels in U-Net and ResNet Not enough timesteps to learn temporal dynamics
- Data parallelism, eventually could use pipeline parallelism for network depth and model parallelism for HiRes data
- Evaluation: RMSE as the optimization target, visualization
- Comparison with linear regression on the full ensemble spread at time t 0h
- Model using only one unperturbed trajectory provides a better spread estimation than using four perturbed trajectories, then 5 ensembles (half of what is available) IPs concatenated to input data
- No improvement using temporal models ENS10, model approximates larger ensembles using only a few input trajectories
- #PAPER Ensemble size: How suboptimal is less than infinity? (Leutbecher 2019)
- #PAPER Multi-model skill assessment of seasonal temperature and precipitation forecasts over Europe (Mishra 2019)
- #PAPER Gronquist 2020 in AI4ES/Bias correction, adjustment
- #TALK
A new approach to subseasonal multi-model forecasting: Online prediction with expert advice (Brayshaw and Gonzalez 2020)
- Tested algorithms to perform âonline prediction with expert adviceâ (Cesa-Bianchi et al. 2006). These methods consider a set of weighted âexpertsâ (usually uniformly weighted at the start of the process) to produce subsequent predictions in which the combination or mixture is updated to optimize a loss or skill function
- S2S4E
- The online learning algorithms
- BOA: Bernstein online aggregation
- MLpol: Polynomial potential aggregation
- Compared to the ’exponentiated gradient’ method as a reference, which is a sequential learning algorithm previously used in weather and climate -> EGA_NWP
- The BOA and MLpol methods show skill improvements for leads beyond week 3, a horizon rarely beaten by ECMWF at the country level
- #PAPER
Multi-model ensemble predictions of precipitation and temperature using machine learning algorithms (Ahmed 2020)
- Optimum performance of multi-model ensemble is achieved with 50% of top-ranked GCMs
- K-Nearest Neighbour and Relevance Vector Machine are good for multi-model ensembles
- Artificial Neural Network multi-model ensembles showed large performance fluctuations in space
- Machine learning-based multi-model ensembles outperformed simple ensemble mean
- #PAPER
A data-driven multi-model ensemble for deterministic and probabilistic precipitation forecasting at seasonal scale (Xu 2020)
- Current numerical models have large uncertainty in model structure, parameterization and initial conditions
- A data-driven multi-model ensemble is constructed using a series of statistical and machine learning methods with varying inputs
- Deterministic precipitation forecasts are produced by the weighting of ensemble members using Bayesian model averaging (BMA) and probabilistic forecasts are generated by sampling from BMA predictive probability density function (PDF)
- The results demonstrate that the accuracy in the statistical ensemble is significantly higher than the North American multi-model ensemble (NMME) for both deterministic and probabilistic precipitation forecasts, especially at 1-month lead
- #PAPER
How to create an operational multi-model of seasonal forecasts? (Hemri 2020)
- #BSC Paco Doblas-Reyes