ARTÍCULO ORIGINAL

DOI: 10.17533/udea.redin.n75a18

Daily rainfall interpolation models obtained by means of genetic programming

Modelos de interpolación de precipitación diaria obtenidos a partir de programación genética

¹ Instituto de Ingeniería, Facultad de Ingeniería, Universidad Nacional Autónoma de México. Av. Universidad 3000, Ciudad Universitaria. México, D.F., México.

² Subcoordinación de Hidrología Superficial, Instituto Mexicano de Tecnología del Agua. Paseo Cuauhnáhuac 8532, Col. Progreso. Morelos, México.

³ Instituto de Investigaciones en Matemáticas Aplicadas y en Sistemas, Universidad Nacional Autónoma de México. Av. Universidad 3000. México, D.F., México.

* Corresponding author: Maritza Liliana Arganis Juárez, e-mail: marganisj@iingen.unam.mx

DOI: 10.17533/udea.redin.n75a18

(Received January 20, 2015; accepted April 20, 2015)

The evolutionary computing algorithm of genetic programming (GP) was applied to obtain mathematical daily rainfall interpolation models in one climatologic station, using the measured data in nearby stations in Cutzamala River basin in Mexico. The obtained models take into account both the geographical coordinates of the climatologic station and also the elevation; the answer of these models was compared against those obtained by means of multiple linear regression, giving genetic programming models a better performance with respect to the multiple linear regression. Isohyets maps were then obtained to compare the spatial shapes between measured and calculated rainfall data in Cutzamala River Basin for a maximum historic storm recorded in year 2006, showing an adequate agreement of the results in case of rainfalls greater than 23 mm. Genetic programming represents a useful practical tool for approaching mathematical models of variables applied in engineering problems and new models could be obtained in several basins by applying these algorithms.

Keywords: daily rainfall, genetic programming, interpolation models, isohyet, geographic coordinates

Resumen

Se aplicó el algoritmo de cómputo evolutivo de programación genética (PG) para obtener modelos matemáticos de interpolación de precipitación diaria en una estación climatológica, utilizando datos medidos en las estaciones cercanas a la cuenca del Río Cutzamala en México. Los modelos obtenidos toman en cuenta tanto las coordenadas geográficas de las estaciones climatológicas como su elevación; la respuesta de los modelos se comparó contra los resultados obtenidos con ayuda de regresiones lineales múltiples, presentando un mejor desempeño programación genética. Adicionalmente, se construyeron mapas de isoyetas para comparar las formas espaciales entre los datos de precipitación medidos y calculados en la cuenca del Río Cutzamala para una tormenta máxima histórica registrada en el año 2006, observándose concordancia en los resultados en el caso de precipitaciones mayores de 23 mm. La programación genética representa una herramienta de utilidad práctica para aproximar modelos matemáticos de variables aplicadas en problemas de ingeniería y se pueden obtener nuevos modelos en distintas cuencas al aplicar estos algoritmos.

Palabras clave: precipitación diaria, programación genética, modelos de interpolación, isoyetas, coordenadas geográficas

Introduction

Precipitation records in many basins of Mexico have a lot of missing data: different times and different existing databases in government agencies (National Water Commission, National Weather Service, the Federal Electricity Commission). There are various techniques of interpolation of precipitation trying to solve this problem such as the method of inverse distance weighted [1], Kriging and Co Kriging methods [2, 3], multifractal analysis and neural networks [4, 5], among others; lots of procedures for fitting curves have been applied in the hydrological field [6-9]. In recent years, evolutionary computation algorithms have been used in hydrology and hydraulics problems to approximate some processes such as the rainfall-runoff relationship, parameters determination on distribution functions, estimation of bivariate distribution functions or approximate equations for estimating the channel temperature from other measured climatological data [8, 10-13].

We use Genetic Programming (GP) algorithm in this paper to approximate the total rainfall in a basin, having information such as geographical coordinates and altitude. We obtained our models with total precipitation data from stations that recorded the storm of July 29^th, 2006 in Cutzamala River Basin; In a first trial we only used 7 from 15 stations in order to find out the GP interpolation model capacity; then, we used 15 stations in the rest of the test. All models were tested in an artificially removed station. Additionally, isohyets from real and calculated data with GP were drawn with Kriging method to compare their shapes. A Co Kriging type method was also used in the removed station. The results obtained with the last GP model were encouraging, as described herein.

Methods

Genetic programming algorithm (GP) [14-16] is a subclass of genetic algorithms; they are inspired in Darwin's Natural Selection Theory, where the best individuals survive from each generation and the remaining disappear. The GP algorithm traditionally consists in randomly generating an initial population of N tree individuals that, in this case, represent mathematical models, formed from a set of primitives (functions and variables) according to the problem to be solved (see Figure 1). Subsequently, as in the case of genetic algorithms [17], each individual is tested in an objective function to evaluate its performance; the best ones are selected (with random methods such as Roulette Wheel Method, Stochastic Universal Method, Tournament). These selected individuals are recombined and mutated in order to generate a new population of N individuals, which becomes the next generation and the process is repeated until a tolerance or a number of generations is reached.

In this study we considered arithmetic operators {+, -, *, /}, transcendent operators {exp, sen, cos} and variables {x, y, z} representing the values ​​of the longitude, latitude in UTM and altitude in thousands of meters, respectively, to obtain a complete precipitation function hp (x, y, z) in mm. A Genetic programming algorithm coded in Matlab [18] and developed in the Research Institute of Applied Mathematics and Systems, National Autonomous University of Mexico was used in the tests performed.

In the first test, we considered 200 individuals and 35 nodes while in the last test we considered 600 individuals and 32 nodes. In all cases, 5000 generations or iterations were set to end the process.

The parameters crossover (recombination) and mutation probabilities of the GP algorithm were set as 90% and 10%, respectively. The considered objective function was defined as the minimization of the mean square error between the function of total rainfall that historically occurred in a selected date and the function of precipitation calculated with genetic programming algorithm (Eq. 1):

Study site

Cutzamala River Basin

Cutzamala River Basin is located in central Mexico; it begins at the birth of Zitacuaro River and the downstream runoff volcanic axis of western slopes of the mountains of Angangueo, Zitacuaro River, which originates in the State of Michoacan, and Tilostoc River in Mexico State, from the ranging station of El Gallo located at the geographical coordinates 100º40'52 ' west longitude and 18º41'15 '' N latitude.

In the 18^th Hydrologic Region Balsas, the annual rainfall volume reaches the 108 370 mm³ , with an annual average rainfall of 927 mm, varying between 873 mm in the Lower Balsas and 1 019 mm in the Medium Balsas. The highest rainfall is in Southern Mother Mountain in Oaxaca and Guerrero with 2 000 mm, and the lower is in Apatzingan Valley in the ''Tierra Caliente michoacana'' with 600 mm. The annual average evaporation reaches 1 750 mm. The rainfalls mainly take place between June and September. The main climate in the region is semi-warm and sub-wet with an annual average temperature of 22° [19].

Input data

The historical storm of July 29^th, 2006, which is one of the greatest historical storms throughout the record, was selected to perform this study.  Rainfall gauge stations with pluviometer, which recorded the height of precipitation (hp) in mm, were chosen on that date; longitude and latitude coordinates (x, y) in UTM (in thousand meters) and altitude (z) in thousands of meters were used to obtain the model. Additionally, the assumption that the weather station 15046 Colorines Dam was not able to measure precipitation was considered. (The station was randomly selected from the cluster of stations allocated approximately in the middle of the basin (Figure 3). This station was not considered for the model (Table 1) and was subsequently used to verify the response model for interpolating missing data.

Results and discussion

In order to find out if with few data the model can be capable to interpolate surrounding data, we applied genetic programming algorithm to the first 7 data from a total of 15 measured precipitation records which appear in Table 2 (in this part we also did not consider the Colorines dam station 15046): Thus, the following mathematical model was determined (Eq. 2):

It was observed the model reported similar results in about 9 of the 15 selected stations, with a minimum absolute error of 0.35 mm in the calculated hp for 15036 Ixtlahuaca station and a relative error of 1.42%; while at station 16136 Tzitzio, the maximum difference of 27.3 mm was obtained with a huge relative error of 682.5% .

The measured and calculated total rainfall data were drawn and compared with an identity function (Figure 4), showing the close grouping to the identity nearly 60% of the data with a determination coefficient of 0.8084. Additionally, the determination coefficient estimated to account for the nonlinear model was obtained, this coefficient gave 0.7916 and was obtained with the variance (var) of the measured data (var_hp) and the variance of the error in the estimation of rainfall (var_hperror). That is C_det = (var_hp- var_hperror) /var_hp.

Subsequently, we used our model to estimate the value of the total precipitation at 15046 Colorines Dam station with coordinates x= 375.5569489 10³ m and y=2112.312711 10³ m and a height z= 1.68 103 m, determining a value of hp=6.60 mm. Moreover, the measured data at that station was 8 mm, indicating the absolute error was 1.4 mm and the relative 17.5%. This data was drawn in the Figure 4 to locate its position regarding the identity function.

Additionally, isohyets were drawn using ARCMAP© and Kriging interpolation, taking into account measured and calculated data with the GP model (Figures 5 and 6).

We can see that Figures 5 and 6 have similar shapes only in their Isohyets near to 20 mm (Figure 4), while for other values the model has difficulties in reproducing the historical behavior.

The mean square error in this model was 512.47, and it was observed that 7 points gave an absolute error lower than 10 mm compared to measured (Table 3). Table 3 shows that the least absolute error was 0.58 mm for the precipitation measuring 23 mm and a relative error of 2.52%, whereas the maximum absolute error was 66.1 mm, for the precipitation of 50 mm with a relative error of 132.2%. Based on this model, the rainfall was calculated for Colorines dam station and the measured and calculated data were plotted against the identity function (Figure 7). In this case, the calculated value for Colorines dam station was 12.93 mm against the 8 mm had occurred historically, it represents an absolute error of 4.93 mm when using Eq. (3) with a relative error of 61.63%. It means Eq. (3) gave a larger error than one obtained with Eq. (2); that is attributed to the dispersion values of the 15 data against the dispersion values of the 7 data.

Taking the 15 stations, the multiple linear regression model obtained with supplements Excel spreadsheet © took the following form (Eq. 4):

The determination coefficient between measured and calculated rainfall data was 0.4734 (Figure 8), which is lower than the result obtained with the GP model.

The estimated rainfall for the Colorines dam station by applying the linear regression model of Eq. (4) was 21.45 mm against the real value of 8 mm; that represents an absolute error of 13.45 mm and a 168% relative error. In the case of the interpolation of a missing data, Genetic programming model of Eq. (2) was the optimal relative error of 17.5%.

The new obtained model was (Eq. 5):

Where:

Based on this latest model, the mean square error of the objective function was 119.15; in this case the determination coefficient was 0.8126, that is a correlation coefficient of 0.9014 (Figure 9). Table 4 shows the measured and the calculated values obtained with Eq. (5).

The interpolated value for Colorines dam station model Eq. (5) was 8.22 mm, representing an absolute error of 0.22 mm and a relative error in percent of 2.75% (Figure 9). This result leads to consider Eq. (5) as the best result of all the tests performed for interpolation purposes and it was the second best result in validating data.

The isohyets corresponding to results of Eq. (5) appear in Figure 10. In this case the model reproduces the shapes of isohyets.

Finally, a Co Kriging method used in scientific field was applied to compare all the results in case of Colorines dam station; the applied optimization model was (Eq. 6):

Where:

The parameters obtained in Excel© with Solver© were as shown in Table 5.

Conclusions

Using genetic programming to interpolate rainfall data to a historical maximum storm recorded at Cutzamala River Basin, was determined with the model of Eq. (2), where an adjustment of about 60% of the data considered was achieved; its obtained isohyets were similar to those recorded historically, mainly for precipitation data from 23 mm. With the model of Eq. (2) the smallest relative interpolation error was obtained for Colorines dam station.

Finally, the GP model of Eq. (5) which takes into account the 15 selected stations gave higher correlations than those obtained with the other GP and multiple linear correlation models and also gave the second best interpolation result for Colorines dam station with a relative error of only 2.75% (because the co Kriging method gave the best interpolation in this analyzed event). Thus, in this respect, the GP model of Eq. (5) can be considered of practical use. And with this model the best Isohyets were obtained as well.

It is very desirable for future developments to use average historical precipitation values ​​in the analyzed sites for generating GP interpolation models, as see the convenience of analyzing a larger number of rainfall events in order to improve the results. 

Acknowledgements

We are very thankful to Ing. Pablo Reyes Godínez for their support for formatting this document.

References

1.  C. Liu, F. Chen. ''Estimation of the spatial rainfall distribution using inverse distance weighting (IDW) in the middle of Taiwan''. Paddy and Water Environment. Vol. 10. 2012. pp. 209-222.         [ Links ]

2.  Z. Bargaoui, A. Chebbi. ''Comparison of two kriging interpolation methods Applied to spatio-temporal rainfall''. Journal of Hydrology. Vol. 365. 2009. pp. 56-73.         [ Links ]

3.  N. Memarsadeghi, D. Mount. Cokriging Interpolation. Scholarly paper for Master's degree, Computer Science Department of University of Maryland. Available on: https://www.cs.umd.edu/sites/default/files/scholarly_papers/cokriging_1.pdf%20. Accessed: December 6th, 2014.         [ Links ]

4.  A. Gires, I. Tchiguirinskaia, D. Schertzer, S. Lovejoy. ''Analyses multifractales et spatio-temporelles des précipitations du modèle Méso-NH et des données radar''. Hydrol. Sci. J. Vol 56. 2011. pp 380-396.         [ Links ]

5.  S. Karmakar, M. Kowar, P. Guhathakurta. Spatial Interpolation of Rainfall Variables using Artificial Neural Network. Proceedings of the International Conference on Advances in Computing, Communication and Control (ICAC3'09). Mumbai, India. 2009. pp. 547-552.         [ Links ]

6.  L. Hay, R. Viger, G. McCabe. ''Precipitation interpolation in mountainous regions using multiple linear regression''. Hydrology, Water Resources and Ecology in Headwaters (Proceedings of the HeadWater'98 Conference held at Meran/Merano, Italy). N.° 248. 1998. pp. 33-38.         [ Links ]

7.  Mair, A. Fares. ''Comparison of Rainfall Interpolation Methods in a Mountainous Region of a Tropical Island''. J. Hydrol. Eng. Vol. 16. 2011. pp. 371-383.         [ Links ]

8.  S. Ly, C. Charles, A. Degré. ''Different methods for spatial interpolation of rainfall data for operational hydrology and hydrological modeling at watershed scale: A review''. Biotechnol. Agron. Soc. Environ. Vol. 17. 2013. pp. 392-406.         [ Links ]

9.  J. Rabuñal, J. Dorado, A. Pazos, D. Pereira. ''A new approach to the extraction of ANN rules and to their generalization capacity through GP''. Neural Computation. Vol. 16. 2004. pp. 1483-1523.         [ Links ]

10.  M. Arganis, R. Val, J. Prats, K. Rodriguez, M. Dolz, R. Dominguez. ''Genetic programming and standardization in water temperature modeling''. Advances in Civil Engineering. Vol. 2009. 2009. pp. 1-10.         [ Links ]

11.  R. Dominguez, O. Fuentes, M. Arganis, A. Mendoza. ''Calculation of Double Bivariate Gumbel probability density function via a Genetic Algorithm: Application to Huites dam basin''. Journal of Flood Engineering (IFE). Vol. 21. 2009. pp. 293-300.         [ Links ]

12.  J. Preciado, M. Ocon, M. Arganis. Approximation of bivariate empirical distribution function of the maximum historic inflow avenues to a dam using genetic programming. Proceedings of the XXII Congress of Hydraulics. Acapulco, Mexico. 2012. pp. 7-9.         [ Links ]

13.  K. Rodriguez, M. Arganis, C. Cruickshank, R. Dominguez. ''Rainfall-runoff modeling using genetic programming''. Journal of Hydroinformatics. Vol. 14. 2012. pp. 108-121.         [ Links ]

14.  W. Banzhaf, P. Nordin, R. Keller, F. Francone. Genetic Programming: An Introduction: On the Automatic Evolution of Computer Programs and Its Applications. 1^st ed. Ed. Morgan Kaufmann. San Francisco, USA. 1998. pp. 481.         [ Links ]

15.  N. Cramer. A Representation for the Adaptive Generation of Simple Sequential Programs. Proceedings of the 1^st Int. Conf. on Genetic Algorithms and the Applications. Pittsburg, USA. 1985. pp. 183-187.         [ Links ]

16.  J. Koza. Hierarchical Genetic Algorithms Operating on Populations of Computer Programs. Proceedings of the 11^th Int. Joint Conf. on Artificial Intelligence. Detroit, USA. 1989. pp. 768-774.         [ Links ]