Universidade Federal de Santa Maria
Ci. e Nat., Santa Maria v.42, Special Edition: Micrometeorologia, e15, 2020
DOI:10.5902/2179460X45358
ISSN 2179-460X
Received: 02/06/20 Accepted: 02/06/20 Published: 28/08/20
Special Edition
Neural network for seasonal climate precipitation prediction on the Brazil
Rede neural para previsão climática sazonal de precipitação no Brasil
Juliana Aparecida Anochi I
Haroldo Fraga de Campos Velho II
I Instituto Nacional de Pesquisas Espaciais, Cachoeira Paulista, Brazil. E-mail: juliana.anochi@inpe.br.
II Instituto Nacional de Pesquisas Espaciais, São Jose dos Campos, Brazil. E-mail: haroldo.camposvelho@inpe.br.
RESUMO
Precipitação é o campo meteorológico mais difícil de ser predito. Uma abordagem baseada em rede neural ótima é aplicada para previsão de precipitação para o Brasil. Uma rede neural perceptron de múltiplas camadas (RN-PMC) auto-configurada é usada como ferramenta predição. A topologia da MLP-NN é encontrada resolvendo um problema de otimização pelo algoritmo de colisão de múltiplas partículas (MPCA). Previsões para estações de inverno e verão são mostradas. A previsão neural é avaliada usando dados de reanálise do NCEP/NCAR e dados do satélite GPCP (Global Precipitation Climatology Project -- monthly precipitation dataset).
Palavras-chave: Precipitação; Previsão climática sazonal; Rede neural auto-configurada.
ABSTRACT
Precipitation is the hardest meteorological field to be predicted. An approach based on and optimal neural network is applied for climate precipitation prediction for the Brazil. A self-configurated multi-layer perceptron neural network (MLP-NN) is used as a predictor tool. The MLP-NN topology is found by solving an optimization problem by the Multi-Particle Collision Algorithm (MPCA). Prediction for Summer and Winter seasons are shown. The neural forecasting is evaluated by using the reanalysis data from the NCEP/NCAR and data from satellite GPCP (Global Precipitation Climatology Project -- monthly precipitation dataset).
Keywords: Precipitation; Seasonal climate prediction; Self-configured neural network.
1 Introduction
The climate forecast is a key factor for many social and economic sectors, such as civil defense for disasters, energy production, agricultural, transportion systems, insurance price. Brazil is a country with a great territorial extension, with different precipitation and temperature regimes. The CPTEC-INPE (CPTEC: Center for Weather Prediction and Climate Studies, INPE: National Institute for Space Research) has global and regional models to perform the weather and climate predictions based on numerical integration of partial differential equations. These are complex and sophisticated computer models executed by a supercomputer.
Models for carrying out prediction and climate monitoring using artificial intelligence are already employed in research related to climate precipitation (ANOCHI and CAMPOS VELHO, 2014), hydrology (SOUZA et al., 2010), severe weather (RUIVO et al., 2015), (BABOO et al., 2010), among others.
Brazil presents a strong climate variability, having equatorial, tropical, and sub-tropical climate regions. In the North, there is a rainy equatorial climate, practically without dry season. In the Northeast, the rainy season, with low rainfall rates, is restricted to a few months, characterizing a semi-arid climate and presents higher climatic predictability. The Southeast and Center-West regions are influenced by both tropical and mid-latitude systems, with a well defined dry season in winter and a rainy season in summer with convective rain-fall. Both regions have low predictability due to less dependence on ocean conditions and the wide variety of meteorological systems that affect them. Finally, the Southern region of Brazil, as well as the North region, do not have a rainy or dry seasons well defined. In the South of Brazil, there is approximately an uniform rainfall distribution during all year. But this region is characterised by medium predictability, and due to its latitudinal location, it is more influenced by medium latitude systems, where frontal systems are the main cause of rainfall during the year.
Several climatic regimes motivated (ANOCHI and SILVA, 2009) to develop a climate prediction model for the precipitation field by Neural Networks (NN). The neural model was adapted to the precipitation field for monthly and seasonal forecasting.
Different from our previous studies, where the NN were designed for climate prediction associated at each Brazilian regions, here the NN forecaster is applied for the entire country.
2 Material and Methods
The proposal is focused on the application of neural networks in meteorology. The methodological novelty consists in the definition of an optimal neural network. The optimal network identification is formulated as an optimization problem, solved by the meta-heuristic MPCA - this methodology is called the self-configuring network.
The methodology can be applied without the support from an expert on neural network. For applying the methodology, the user provides a data set, and the system defines the best neural network type multilayer perceptron type for the application.
2.1 Neural Networks
Artificial neural networks are computational methods in which the operating principle is conducted by a mathematical model inspired by the functioning of the basic elements that form the neural structure of intelligent organisms, that acquire knowledge through experience. Their behavior results from the interactions between the processing units, from their environment, through a learning process.
Neural networks are distributed parallel systems, composed of neurons or processing units, which compute certain mathematical functions, usually non-linear. Processing neurons can be distributed in one or more layers and interconnected by a large number of connections (synaptic weights), which store the knowledge represented in the model.
Mathematically, we can describe a neuron k by writing the following pair of equations (HAYKIN, 2001):
|
|
(1) |
|
|
(2) |
where are the inputs;
are the connection
weights;
is the bias;
is the activation function; and
is the output.
The different architectures of neural networks can be formed by the combination of artificial neurons and are defined by the type of connection between networks. Each neuron transfers its signal to the neurons that are in one of the subsequent layers.
In this work, the Multiple Layer Perceptron (MLP) network was used. The MLP has emerged as an alternative solution for non-linearly separable problems and has been successfully applied to solve complex problems through its supervised training using the error backpropagation algorithm based on the learning rule for correction of error (HAYKIN, 2001).
The architecture of the MLP network consists of the topological arrangement of the processing units of the neurons with the respective values of weights associated with the connections. The synaptic weights are adjusted by delta rule. The MLP network has an input layer, at least one intermediate layer, and an output layer.
Although a NN model has great potential, its performance depends on the definition of its parameters, since the definition of the topology can significantly influence the phase of the training process.
2.2 MPCA to identify the best topology NN
In practice, the NN topology is usually selected by using empirical or statistical methods that are used to find the best parameters. Here, we use the Multiple Particle Collision Algorithm (MPCA) metaheuristic for configuration the topology of the MLP network. The strategy applied by ANOCHI and CAMPOS VELHO, 2014, can be considered as an optimization problem, where each point in the search space represents a NN with different topologies.
The MPCA metaheuristic was introduced by (LUZ et al., 2008), and it is an extension of the canonical Particle Collision Algorithm (PCA) (SACCO, 2005). The proposed structure for the MPCA algorithm uses a set of n particles, independently exploring but collaboratively, the same search space. The introduction of n particles leads to the need to implement an indirect communication mechanism between particles.
The MPCA starts with a selection of an initial solution, it is modified by a stochastic perturbation conducing to the construction of a new solution. The new solution is compared and this solution can or cannot be accepted. If the new solution is not accepted, the particle can be send to a different location of the search space. If a new solution is better than the new solution is absorbed. The figure 1 shows the pseudo-code of the MPCA metaheuristics (LUZ, 2008).
Figure 1 – Pseudo-code for MPCA. Adaptada de (LUZ, 2008)
The MPCA has been used successfully in several optimization problems, such as: diagnosis of failures (ECHEVARRÍA et al., 2014), identification of atmospheric temperature profile (SAMBATTI et al., 2012), climate prediction (ANOCHI and CAMPOS VELHO, 2014), solution of a radiative inverse problem (TORRES et al., 2015), and the other applications.
2.3 Meteorological data
For the climatic forecasting of precipitation using neural networks, different data sets were collected, from reanalysis data from the National Centers for Environmental Prediction (NCEP/NCAR) to Global Precipitation Climatology Project monthly precipitation dataset (GPCP).
The NCEP/NCAR provides historical series of meteorological data obtained through the assimilation and analysis of data observed for the entire planet from 1948 up to the present. These data come from radiosondes, land surface meteorological stations, oceanic buoys, ship, airplanes, satellites, GNSS (General Navigation Satellite Systems) stations, and data from model (control) and analysis (assimilation process). For the constitution of these data, global atmospheric and surface flow fields derived from numerical forecasting and data assimilation systems are used (KALNEY, 1996).
The input variables to the neural network were selected from the NCEP Reanalysis data from the NOAA/OAR/ESRL PSD, Boulder (CO), USA, by the website: {\tt https://www.esrl.noaa.gov/psd/}. The input variables are: zonal and meridional winds at three levels (300hPa, 500hPa, 850hPa), air temperature at 850hPA, and specific humidity at 850hPa.
The GPCP Monthly product provides a consistent analysis of global precipitation from an integration of various satellite data sets over land and ocean, and a analysis calibration over land. Data from rain gauge stations, satellites, and sounding observations have been merged to estimate monthly rainfall on a 2.5º global grid from 1979 to the present (ADLER, 2003).
The variable used as output to the NN was selected from the monthly precipitation dataset - GPCP provided by the NOAA/OAR/ESRL PSD, Colorado, USA.
3 Experimental settings
The strategy for optimizing the NN architecture
is considered as a mono-objective optimization problem. Four search space
parameters will be optimized: two continuous variables (the learning rate
parameter , and the momentum
constant (
, and two discrete
variables (the number of neurons in the hidden layer, and the type of
activation function).
For the neural network learning phase, a well
known procedure is the delta rule, where a correction is applied to the
synaptic weight, for minimizing the square difference between the network
output and the target values
. Following the rule
delta, the synaptic weight are found by (HAYKIN, 2001):
|
|
(3) |
where -
.
In this experiment, several realizations are executed to find representative solutions. The configuration of the NN topology for the climatic prediction experiment is described below - the NN is applied for each grid point:
· Nine inputs (meteorological variables).
·
One output node for the results: the
precipitation at the grid point. In the training algorithm, the MLP computes
the output and does a comparison
with the target data
(observed
precipitation).
· One hidden layer with 12 neurons.
· The hyperbolic tangent is the activation function.
·
Learning rate and momentum value
produce the best
fitness (MLP-MPCA) for the following numerical values:
and
.
·
The iteration the training phase stops when the
error reaches the value .
· For determining the NN configuration by the MPCA, 25 experiments are performed to find the best fitness to NN.
4 Results and Discussion
The results presented in this section show the behavior of the neural network as a predictor model of the precipitation variable.
The numerical experiments carried out in this study show a good performance of using the MPCA tool. The procedure identifies the best parameters for the NN application, finding its topology by minimizing the a cost functional. The mentioned approach does not require a specialist to the task for NN configuration. The neural model can be implemented for both applications, for operational use or/and in research activities.
Figure 2(a) is the observed precipitation by GPCP in the summer at 2015 in Brazil. Figure 2(b) is the forecast of precipitation obtained by the Artificial Neural Network (ANN) at same period (summer 2015). The NN presented a good performance for describing the precipitation patterns for the Brazilian regions: North, Northeast, Midwest, and Southeast. In the states of São Paulo and Mato Grosso do Sul, the NN underestimated the precipitation. For the Santa Catarina state, the NN model overestimated the precipitation. However, in general, the results obtained with the neural model showed a good behavior when compared with the observed data (GPCP).
Figure 2 – Climate precipitation prediction in Brazil for summer 2015
Figure 3(a) is the precipitation observed by GPCP in the autumn at 2015 in Brazil. Figure 3(b) is the prediction of precipitation predicted by ANN in the autumn 2015. For the cited season, the neural network did a good prediction for the rainfall behavior for the regions: in the east of the northeast (period of the rainy season); the extreme north of Brazil, and in the South Region of Brazil, composed by the states of Paraná, Santa Catarina, and Rio Grande do Sul - these three states form the South (Brazilian) region, presenting a subtropical climate regime, and the precipitation is well distributed throughout the year in this region.
Figure 3 – Climate precipitation prediction in Brazil for autumn 2015
Figure 4(a) is the precipitation observed by GPCP in the winter at 2015 in Brazil. Figure 4(b) is the prediction obtained with the neural model. The forecast of the neural model (ANN-MPCA) for winter presented a pattern very similar to the observation of the GPCP. The exception occurred in northern Maranhão, in eastern Piauí, in the states of Ceará, Rio Grande do Norte, Paraíba and Pernambuco, where the neural model overestimated precipitation.
Figure 4 – Climate precipitation prediction in Brazil for winter 2015
Figure 5(a) is the precipitation observed by GPCP in the Spring at 2015 in Brazil. Figure 5(b) is the prediction obtained with the neural model. During this period of spring, the neural model shows a good forecast in almost the entire Brazilian territory, mainly in the north, northeast and central west regions.
Figure 5 – Climate precipitation prediction in Brazil for Spring 2015
5 Conclusion
Brazil is a country with different climatic conditions: equatorial, tropical, and subtropical climatic zones. The variable precipitation has a high variability implying in difficulties to be predicted. Precipitation has a strong impact for the society (natural disasters) and for several economical sectors.
It is a hard task to develop models to predict the precipitation. There are many reasons for such difficulties: local and synoptic patterns has different influences, and different variables and/or conditions are associated with the rain-fall depending on the season of the year. In addition, the precipitation has high variability in space and time. All these factors has strongly influence in the behavior of the precipitation.
Neural network is a prestigious area of Artificial Intelligence and has shown their efficiency in various application areas, being meteorology one of them. Predictive models based on NNs are easy to use, and do not require very powerful computers.
The results present in this paper are consistent with observation data (GPCP), becoming a good tool to support seasonal climate prediction.
Acknowledgments
The author J. A. Anochi thanks by the use of the supercomputer facilities from the CPTEC-INPE. Haroldo F. Campos Velho thanks CNPq and FAPESP.
References
ADLER, R. F.; HUFFMAN, G. J.; CHANG, A.; FERRARO, R.; XIE, P.-P.; JANOWIAK, J.; RUDOLF, B.; SCHNEIDER, U.; CURTIS, S.; BOLVIN, D. et al. The version-2 global precipitation climatology project (GPCP) monthly precipitation analysis (1979–present). Journal of hydrometeorology, v. 4, n. 6, p. 1147-1167, 2003.
ANOCHI, J.; SAMBATTI, S.; LUZ, E.; CAMPOS VELHO, H. F. New learning strategy for supervised neural network: MPCA meta-heuristic approach. In: 1st BRICS Countries & 11th CBIC Brazilian Congress on Computational Intelligence. Location: Recife, Brasil. Porto de Galinhas Beach. 2013.
ANOCHI, J.; SILVA, J. Uso de redes neurais artificiais e teoria de conjuntos aproximativos no estudo de padrões climáticos sazonais. Learning and Nonlinear Models, v. 7, p. 83–91, 2009.
ANOCHI, J. A.; CAMPOS VELHO, H. F. Optimization of feedforward neural network by multiple particle collision algorithm. In: IEEE. 2014 IEEE Symposium on Foundations of Computational Intelligence (FOCI), p. 128–134, 2014.
BABOO, S. S.; SHEREEF, I. K. An efficient weather forecasting system using artificial neural network. International journal of environmental science and development, v. 1, n. 4, p. 321, 2010.
ECHEVARRÍA, L. C.; SANTIAGO, O. L.; NETO, A. J.S. Aplicación de los algoritmos evolución diferencial y colisión de partículas al diagnóstico de fallos en sistemas industriales. Investigación Operacional, v. 33, n. 2, p. 160–172, 2014.
HAYKIN, S. Neural networks: a comprehensive foundation. Prentice Hall PTR, 1994.
KALNAY, E.; KANAMITSU, M.; KISTLER, R.; COLLINS, W.; DEAVEN, D.; GANDIN, L.; IREDELL, M.; SAHA, S.; WHITE, G.; WOOLLEN, J. et al. The NCEP/NCAR 40-year reanalysis project. Bulletin of the American meteorological Society, American Meteorological Society, v. 77, n. 3, p. 437–472, 1996.
LUZ, E. F. P. da; BECCENERI, J. C.; CAMPOS VELHO, H. F. A new multi-particle collision algorithm for optimization in a high performance environment. Journal of Computational Interdisciplinary Sciences, v. 1, n. 1, p. 3–10, 2008.
RUIVO, H. M.; CAMPOS VELHO, H. F. de; SAMPAIO, G.; RAMOS, F. M. Analysis of extreme precipitation events using a novel data mining approach. American Journal of Environmental Engineering, v. 5, n. 1A, p. 96–105, 2015.
SACCO, W. F.; OLIVEIRA, C. R. D. A new stochastic optimization algorithm based on a particle collision metaheuristic. Proceedings of 6th WCSMO, 2005.
SAMBATTI, S. B. M.; ANOCHI, J. A.; LUZ, E. F. P.; CARVALHO, A. R.; SHIGUEMORI, E. H.; CAMPOS VELHO, H. C. Automatic configuration for neural network applied to atmospheric temperature profile identification. In: 3rd International Conference on International Conference on Engineering Optimization, p. 1–9, 2012.
SOUSA, W. d. S.; SOUSA, F. d. A. de. Rede neural artificial aplicada à previsão de vazão da bacia hidrográfica do rio piancó. Revista Brasileira de Engenharia Agricola e Ambiental-Agriambi, v. 14, n. 2, 2010.
TORRES, R. H.; LUZ, E.; CAMPOS VELHO, H. C. Multi-particle collision algorithm for solving an inverse radiative problem. In: Integral Methods in Science and Engineering. Birkhäuser, Cham, 2015. p. 309-319.