scholarly journals Testing the performance of three nonlinear methods of time seriesanalysis for prediction and downscaling of European daily temperatures

2005 ◽  
Vol 12 (6) ◽  
pp. 979-991 ◽  
Author(s):  
J. Miksovsky ◽  
A. Raidl
Keyword(s):  
Neural Network ◽  
Linear Regression ◽  
Large Scale ◽  
Linear Models ◽  
Transfer Functions ◽  
Reanalysis Data ◽  
Nonlinear Method ◽  
Nonlinear Methods ◽  
Nonlinear Character ◽  
Short Term Prediction

Abstract. We investigated the usability of the method of local linear models (LLM), multilayer perceptron neural network (MLP NN) and radial basis function neural network (RBF NN) for the construction of temporal and spatial transfer functions between different meteorological quantities, and compared the obtained results both mutually and to the results of multiple linear regression (MLR). The tested methods were applied for the short-term prediction of daily mean temperatures and for the downscaling of NCEP/NCAR reanalysis data, using series of daily mean, minimum and maximum temperatures from 25 European stations as predictands. None of the tested nonlinear methods was recognized to be distinctly superior to the others, but all nonlinear techniques proved to be better than linear regression in the majority of the cases. It is also discussed that the most frequently used nonlinear method, the MLP neural network, may not be the best choice for processing the climatic time series - LLM method or RBF NNs can offer a comparable or slightly better performance and they do not suffer from some of the practical disadvantages of MLPs. Aside from comparing the performance of different methods, we paid attention to geographical and seasonal variations of the results. The forecasting results showed that the nonlinear character of relations between climate variables is well apparent over most of Europe, in contrast to rather weak nonlinearity in the Mediterranean and North Africa. No clear large-scale geographical structure of nonlinearity was identified in the case of downscaling. Nonlinearity also seems to be noticeably stronger in winter than in summer in most locations, for both forecasting and downscaling.

scholarly journals Modeling Nonlinear Transfer Functions from Speech Envelopes to Encephalography with Neural Networks

2019 ◽  
Vol 11 (4) ◽  
pp. 1 ◽  
Author(s):  
Tobias de Taillez ◽  
Florian Denk ◽  
Bojana Mirkovic ◽  
Birger Kollmeier ◽  
Bernd T. Meyer
Keyword(s):  
Neural Network ◽  
Neural Networks ◽  
Artificial Neural Network ◽  
Linear Models ◽  
Transfer Functions ◽  
Event Related Potentials ◽  
Stimulus Presentation ◽  
Related Potentials ◽  
Linear And Nonlinear ◽  
Artificial Neural

Diferent linear models have been proposed to establish a link between an auditory stimulus and the neurophysiological response obtained through electroencephalography (EEG). We investigate if non-linear mappings can be modeled with deep neural networks trained on continuous speech envelopes and EEG data obtained in an auditory attention two-speaker scenario. An artificial neural network was trained to predict the EEG response related to the attended and unattended speech envelopes. After training, the properties of the DNN-based model are analyzed by measuring the transfer function between input envelopes and predicted EEG signals by using click-like stimuli and frequency sweeps as input patterns. Using sweep responses allows to separate the linear and nonlinear response components also with respect to attention. The responses from the model trained on normal speech resemble event-related potentials despite the fact that the DNN was not trained to reproduce such patterns. These responses are modulated by attention, since we obtain significantly lower amplitudes at latencies of 110 ms, 170 ms and 300 ms after stimulus presentation for unattended processing in contrast to the attended. The comparison of linear and nonlinear components indicates that the largest contribution arises from linear processing (75%), while the remaining 25% are attributed to nonlinear processes in the model. Further, a spectral analysis showed a stronger 5 Hz component in modeled EEG for attended in contrast to unattended predictions. The results indicate that the artificial neural network produces responses consistent with recent findings and presents a new approach for quantifying the model properties.

scholarly journals An adaptive two-stage analog/regression model for probabilistic prediction of small-scale precipitation in France

2018 ◽  
Vol 22 (1) ◽  
pp. 265-286 ◽  
Author(s):  
Jérémy Chardon ◽  
Benoit Hingray ◽  
Anne-Catherine Favre
Keyword(s):  
Regression Model ◽  
Large Scale ◽  
Linear Models ◽  
Transfer Functions ◽  
Small Scale ◽  
Probabilistic Prediction ◽  
Two Stage ◽  
Precipitation Occurrence ◽  
Seasonal Stratification ◽  
Local Weather

Abstract. Statistical downscaling models (SDMs) are often used to produce local weather scenarios from large-scale atmospheric information. SDMs include transfer functions which are based on a statistical link identified from observations between local weather and a set of large-scale predictors. As physical processes driving surface weather vary in time, the most relevant predictors and the regression link are likely to vary in time too. This is well known for precipitation for instance and the link is thus often estimated after some seasonal stratification of the data. In this study, we present a two-stage analog/regression model where the regression link is estimated from atmospheric analogs of the current prediction day. Atmospheric analogs are identified from fields of geopotential heights at 1000 and 500 hPa. For the regression stage, two generalized linear models are further used to model the probability of precipitation occurrence and the distribution of non-zero precipitation amounts, respectively. The two-stage model is evaluated for the probabilistic prediction of small-scale precipitation over France. It noticeably improves the skill of the prediction for both precipitation occurrence and amount. As the analog days vary from one prediction day to another, the atmospheric predictors selected in the regression stage and the value of the corresponding regression coefficients can vary from one prediction day to another. The model allows thus for a day-to-day adaptive and tailored downscaling. It can also reveal specific predictors for peculiar and non-frequent weather configurations.

scholarly journals An adaptive two-stage analog/regression model for probabilistic prediction of local precipitation in France

2017 ◽  
Author(s):  
Jérémy Chardon ◽  
Benoit Hingray ◽  
Anne-Catherine Favre
Keyword(s):  
Hybrid Model ◽  
Large Scale ◽  
Linear Models ◽  
Transfer Functions ◽  
Hybrid Approach ◽  
Probabilistic Prediction ◽  
Precipitation Occurrence ◽  
Seasonal Stratification ◽  
Probability Of Precipitation ◽  
Local Weather

Abstract. Statistical Downscaling Methods (SDMs) are often used to produce local weather scenarios from large scale atmospheric information. SDMs include transfer functions which are based on a statistical link identified from observations between local weather and a set of large scale predictors. As physical processes generating surface weather vary in time, the most relevant predictors and the regression link are likely to also vary in time. This is well known for precipitation for instance and the link is thus often estimated after some seasonal stratification of the data. In this study, we present a hybrid model where the regression link is estimated from atmospheric analogs of the current prediction day. Atmospheric analogs are first identified from geopotential fields at 1000 and 500 hPa. For the regression stage, two Generalized Linear Models are further used to model the probability of precipitation occurrence and the distribution of non-zero precipitation amounts respectively. The hybrid model is evaluated for the probabilistic prediction of local precipitation over France. It noticeably improves the skill of the prediction for both precipitation occurrence and quantity. As the analog days vary from one prediction day to another, the atmospheric predictors selected in the regression stage and the value of the corresponding regression coefficients vary from one prediction day to another. The hybrid approach allows thus for a day-to-day adaptive and tailored downscaling. It can also reveal specific predictors for peculiar and non-frequent weather configurations.

BP Neural Network and Multiple Linear Regression in Acute Hospitalization Costs in the Comparative Study

2011 ◽  
Vol 50-51 ◽  
pp. 959-963 ◽  
Author(s):  
Jian Hui Wu ◽  
Guo Li Wang ◽  
Jing Wang ◽  
Yu Su
Keyword(s):  
Neural Network ◽  
Linear Regression ◽  
Bp Neural Network ◽  
Large Scale ◽  
Nonlinear Dynamic System ◽  
Hospitalization Costs ◽  
Self Adaptation ◽  
Material Factor ◽  
The Comparative Study ◽  
Self Learning

The BP neural network is the important component of artificial neural networks, and gradually becomes a branch of the computation statistics. With its many characteristics such as large-scale parallel information processing, excellent self-adaptation and self-learning, the BP neural network has been used in solving the complex nonlinear dynamic system prediction. The BP neural network does not need the precise mathematical model, does not have any supposition request to the material itself. Its processing non-linear problem's ability is stronger than traditional statistical methods. By means of contrasting the BP neural network and the multi-dimensional linear regression ,this article discoveries that the BP neural network fitting ability is more stronger, the prediction performance is more stable, may be further applied and promoted in analysis and forecast of the continual material factor.

scholarly journals Differential Credibility Assessment for Statistical Downscaling

2020 ◽  
Vol 59 (8) ◽  
pp. 1333-1349
Author(s):  
S. C. Pryor ◽  
J. T. Schoof
Keyword(s):  
Large Scale ◽  
Climate Models ◽  
Linear Models ◽  
Transfer Functions ◽  
Global Climate ◽  
Global Climate Models ◽  
Maximum Temperature ◽  
Climate Projections ◽  
Climate Extreme ◽  
Credibility Assessment

AbstractClimate science is increasingly using (i) ensembles of climate projections from multiple models derived using different assumptions and/or scenarios and (ii) process-oriented diagnostics of model fidelity. Efforts to assign differential credibility to projections and/or models are also rapidly advancing. A framework to quantify and depict the credibility of statistically downscaled model output is presented and demonstrated. The approach employs transfer functions in the form of robust and resilient generalized linear models applied to downscale daily minimum and maximum temperature anomalies at 10 locations using predictors drawn from ERA-Interim reanalysis and two global climate models (GCM; GFDL-ESM2M and MPI-ESM-LR). The downscaled time series are used to derive several impact-relevant Climate Extreme (CLIMDEX) temperature indices that are assigned credibility based on 1) the reproduction of relevant large-scale predictors by the GCMs (i.e., fraction of regression beta weights derived from predictors that are well reproduced) and 2) the degree of variance in the observations reproduced in the downscaled series following application of a new variance inflation technique. Credibility of the downscaled predictands varies across locations and between the two GCM and is generally higher for minimum temperature than for maximum temperature. The differential credibility assessment framework demonstrated here is easy to use and flexible. It can be applied as is to inform decision-makers about projection confidence and/or can be extended to include other components of the transfer functions, and/or used to weight members of a statistically downscaled ensemble.

Global assessment of linear and non-linear statistical-dynamical hindcast models of Tropical Cyclones intensity

2020 ◽  
Author(s):  
Neetu Suresh ◽  
Lengaigne Matthieu ◽  
Vialard Jerome ◽  
Mangeas Morgan ◽  
Menkes Christophe ◽  
...  
Keyword(s):  
Linear Regression ◽  
Tropical Cyclones ◽  
Large Scale ◽  
Linear Models ◽  
Environmental Parameters ◽  
Rate Of Change ◽  
Statistical Prediction ◽  
Support Vector ◽  
Environmental Predictors ◽  
Non Linear

<p>Tropical cyclones (hereafter TC) are amongst the most devastating natural phenomena for coastal regions worldwide. While there has been tremendous progress in forecasting TC tracks, intensity forecasts have been trailing behind. Most operational statistical-dynamical forecasts of TC intensity use linear regression techniques to relate the initial TC characteristics and most relevant large-scale environmental parameters along the TC track to the TC intensification rate. Historically, different operational prediction schemes have been developed independently for each TC-prone basin, making it difficult to compare skills between different TC basins. We have thus developed global TC intensity hindcasts using consistent predictors derived from a single atmospheric dataset over the same period. Linear hindcast schemes were built separately for each TC basin, based on multiple linear regression. They display comparable skill to previously-described similar hindcast schemes, and beat persistence by 20–40% in most basins, except in the North Atlantic and northern Indian Ocean, where the skill gain is only 10–25%. Most (60–80%) of the skill gain arises from the TC characteristics (intensity and its rate of change) at the beginning of the forecast, with a relative contribution from each environmental parameter that is strongly basin-dependent. Hindcast models built from climatological environmental predictors perform almost as well as using real-time values, which may allow to considerably simplify operational implementation in such models. Our results finally reveal that these models have 2 to 4 times less skill in hindcasting moderate (Category 2 and weaker) than in hindcasting strong TCs.</p><p>This last result suggests that linear models may not be sufficient for TC intensity hindcasts. Many physical processes involved in TCs intensification are indeed non-linear. We hence further investigated the benefits of non-linear statistical prediction schemes, using the same set of input parameters as for the linear models above. These schemes are based on either support vector machine (SVM) or artificial neural network algorithms. Contrary to linear schemes, which perform slightly better when trained individually over each TC basin, non-linear methods perform best when trained globally. Non-linear schemes improve TC intensity hindcasts relative to linear schemes in all TC-prone basins, especially SVM for which this improvement reaches ~10% globally, partly because they better use the non-seasonal variations of environmental predictors. The SVM scheme, in particular, partially corrects the tendency of the linear scheme to underperform for Category 2 and weaker TCs. Although the TC intensity hindcast skill improvements described above are an upper limit of what could be achieved operationally, it is comparable to that achieved by operational forecasts over the last 20 years. This improvement is sufficiently large to motivate more testing of non-linear methods for statistical TC intensity prediction at operational centres.</p>

scholarly journals A neural-network framework for modelling auditory sensory cells and synapses

2020 ◽  
Author(s):  
Fotios Drakopoulos ◽  
Deepak Baby ◽  
Sarah Verhulst
Keyword(s):  
Neural Network ◽  
Computational Neuroscience ◽  
Large Scale ◽  
Transfer Functions ◽  
Analytical Models ◽  
Sensory Cells ◽  
Ann Model ◽  
Nonlinear Properties ◽  
Model Training ◽  
Artificial Neural Network Ann

AbstractIn classical computational neuroscience, transfer functions are derived from neuronal recordings to derive analytical model descriptions of neuronal processes. This approach has resulted in a variety of Hodgkin-Huxley-type neuronal models, or multi-compartment synapse models, that accurately mimic neuronal recordings and have transformed the neuroscience field. However, these analytical models are typically slow to compute due to their inclusion of dynamic and nonlinear properties of the underlying biological system. This drawback limits the extent to which these models can be integrated within large-scale neuronal simulation frameworks and hinders an uptake by the neuro-engineering field which requires fast and efficient model descriptions. To bridge this translational gap, we present a hybrid, machine-learning and computational-neuroscience approach that transforms analytical sensory neuron and synapse models into artificial-neural-network (ANN) neuronal units with the same biophysical properties. Our ANN-model architecture comprises parallel and differentiable equations that can be used for backpropagation in neuro-engineering applications, and offers a simulation run-time improvement factor of 70 and 280 on CPU or GPU systems respectively. We focussed our development on auditory sensory neurons and synapses, and show that our ANN-model architecture generalizes well to a variety of existing analytical models of different complexity. The model training and development approach we present can easily be modified for other neuron and synapse types to accelerate the development of large-scale brain networks and to open up avenues for ANN-based treatments of the pathological system.

Sign in / Sign up

Export Citation Format

Share Document