1 / 26

Applying a Dynamic Data Driven Genetic Algorithm to Improve Forest Fire Spread Prediction

Applying a Dynamic Data Driven Genetic Algorithm to Improve Forest Fire Spread Prediction. Mónica Denham, Ana Cortés, Tomàs Margalef and Emilio Luque. monica@caos.uab.es, {ana.cortes,tomas.margalef,emilio.luque}@uab.es. Computer Architecture & Operating Systems Department. Content.

toni
Download Presentation

Applying a Dynamic Data Driven Genetic Algorithm to Improve Forest Fire Spread Prediction

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Applying a Dynamic Data Driven Genetic Algorithm to Improve Forest Fire Spread Prediction Mónica Denham, Ana Cortés, Tomàs Margalef and Emilio Luque monica@caos.uab.es, {ana.cortes,tomas.margalef,emilio.luque}@uab.es Computer Architecture & Operating Systems Department

  2. Content • Problem. • Method description. • Dynamic Data Driven Genetic Algorithm: • Analytical Method. • Experiments and Results. • Conclusions.

  3. Problem Initial fire Fire behavior Simulator Terrain slope Terrain vegetation Wind direction Wind velocity Live fuel moisture Dead fuel moisture Uncertainty source

  4. Content • Problem. • Method description. • Dynamic Data Driven Genetic Algorithm. • Analytical Method. • Experiments and Results. • Conclusions.

  5. Method Description (I) Prediction (ti+1) Init (ti) Fire Simulator real spread input parameters Calibration (ti+1) Init (ti) Prediction (ti+2) Fire Simulator Fire Simulator Best set of parameters ¿ input parameters input parameters * real spread

  6. Method Description (II) Calibration (ti+1) Init (ti) Prediction (ti+2) Best set of parameters Fire Simulator Fire Simulator ¿ input parameters input parameters Terrain slope (static) Terrain vegetation (static) Wind direction (dynamic) Wind velocity (dynamic) Live fuel humidity (dynamic) Dead fuel humidity (dynamics) (1h, 10hs, 100hs) real spread master worker0 simulation comparison Population dealing multiple scenarios huge search space Population gathering workern Genetic Operators * Genetic Algorithm simulation comparison

  7. Content • Problem. • Method description. • Dynamic Data Driven Genetic Algorithm: • Analytical Method. • Experiments and Results. • Conclusions.

  8. Dynamic Data Driven Genetic Algorithm (I) DDDAS: ability to dynamically incorporate additional data into an execution application, and in reverse, ability of an application to dynamically steer the measurement process (F. Darema). Improving simulations at Calibration stage and improving overall Prediction process results Analyzing real fire spread (Real fire spread at ti+1) We can Inject better parameter values

  9. Dynamic Data Driven Genetic Algorithm (II) • Due to main fire physics aspects, wind and slope determine fire spread main conditions: fire wind slope Using fire and slope characteristics we can estimate wind values to imitate real fire spread fire • We know slope main characteristics. • Due to Calibration stage requirements we dispose fire spread information from instant ti to ti+1. wind slope We had proposed a method that calculates ideal wind characteristics, with which, in combination with slope aspects, achieves fire propagation like fire spread observed at instant ti+1.

  10. Rw Rw Rs Rs α α β β we know X, Y, we have (Rs y β) Isolating from the equations: We can calculate wind direction and speed (Rw y α ) ( x - Rs cos(β) ) Rw = ( y - Rs sin(β) ) α = arctan cos(α) ( x - Rs cos(β) ) Dynamic Data Driven Genetic Algorithm: Analytical Method (III) Analytical Method: • Simulator uses slope and wind characteristics to calculate fire spread characteristics. Simulator treats slope and wind as vectors. • We use simulator idea to calculate wind characteristics from slope and spread characteristics. X = Rs cos(β) + Rwcos(α) Y = Rs sin(β) + Rwsin(α) Simulator: Real map: x, y Where: Rs : slope factor β: slope direction Rw: wind factor α: wind direction

  11. Dynamic Data Driven Genetic Algorithm (IV) Calibration (ti+1) Init (ti) Fire Simulator input parameters input parameters Terrain slope Terrain vegetation Wind direction Wind velocity Live fuel moisture Dead fuel moisture (1h, 10hs, 100hs) wind max Wind direction Wind velocity slope real spread Genetic Algorithm Selection - Elitism Crossover Mutation *

  12. Content • Problem. • Method description. • Dynamic Data Driven Genetic Algorithm. • Analytical Method. • Experiments and Results. • Conclusions.

  13. Experiments & Results (I) • DDD Genetic Algorithm reduces • error for both Calibration and • Prediction stages. • Prediction stage shows bigger errors • (this stage works with calibration best • individual for previous time step).

  14. Experiments & Results (II) • DDD Genetic Algorithm reduces • error for both Calibration and • Prediction stages.

  15. Experiments & Results (III) • Real map: fire behavior is different • through time steps. • Errors are bigger than previous maps • (synthetic maps). • Errors are similar for different • configurations of our method.

  16. Content • Problem. • Method description. • Dynamic Data Driven Genetic Algorithm. • Analytical Method. • Experimentation and Results. • Conclusions.

  17. Conclusions • Calibration and Prediction stages have shown expected behavior. • We have used these techniques for real and synthetic burnings. Proposed methods shown good performance. • Using DDD Genetic Algorithm we could improve whole prediction process quality for synthetic cases. • Although real fires are our main objective, synthetic cases are the first step for understanding and improving steering methods. • Real fires characteristics are more difficult to simulate. Simulators implement abstractions of reality. We are working in this topic now.

  18. Applying a Dynamic Data Driven Genetic Algorithm to Improve Forest Fire Spread Prediction Thanks monica@aomail.uab.es, {ana.cortes,tomas.margalef,emilio.luque}@uab.es Computer Architecture & Operating Systems Department

  19. Pseudocódigos. • Fórmulas • Fitness - Error / Volver. • Viento / Volver. • Pendiente / Volver.

  20. Pseudocódigo: Algoritmo Genético while (i<=(gen->size -2)) { int p1 =select(totfitc, gen); int p2 =select(totfitc, gen); crossover(gen->gen[p1],gen->gen[p2], &newgen.gen[i], &newgen.gen[i+1]); /*mutation */ mutate(&newgen.gen[i], lmin,lmax); mutate(&newgen.gen[i+1],lmin,lmax); / /* stay in the limites */ clip(&newgen.gen[i]); clip(&newgen.gen[i+1]); // next i = i+2; } void evolute(gentype * gen) { double totfitc; gen->num++; newgen.num = gen->num; cal_fit( gen ,&totfitc); nsort(gen); //keep_best(gen); indcpy(&newgen.gen[0],*get_best()); indcpy(&newgen.gen[1],gen->gen[1]); Volver

  21. Pseudocódigo: Algoritmo Genético (cont.) void crossover(indvtype og1, indvtype og2,indvtype * ng1,indvtype * ng2) { double cros =rand() % 1000; if (cros <= crosp) //si se supera la probabilidad se cruzan { crospoint = (int) rand()% og1.n; for(int i=0;i<crospoint;i++) { ng1->p[i] = og1.p[i]; ng2->p[i] =(og2.p[i]+og1.p[i])/2; } for(int i=crospoint;i<og1.n;i++) { ng1->p[i] =(og2.p[i]+og1.p[i])/2; ng2->p[i] =og2.p[i]; } } else // se copian directamente los padres a los hijos { indcpy(ng1,og1); indcpy(ng2,og2); } } int select( double totfitc,gentype * gen) { double r = random * totfitc; p=0; sumfit=0; while((sumfit <= r)&&(p<gen->size)) sumfit += gen->gen[p++].fitc; p--; return p; } Volver

  22. Viento • Sigma = acumulación de la contribución a la intensidad de reacción (modelo comb.) • Beta = acumulación carga/densidad de todas las partículas del modelo. • betaOpt = 3.348 / sigma0.8189 • Ratio = beta/betaOpt • c = 7.47 * exp(-0.133 * sigma0.55)) • e = 0.715 * exp(-0.000359 * sigma) • WindK = c * ratio-e • WindB = 0.02526 * sigma0.54 • Fuel_PhiWind = WindK * windSpeedWindB • Rw = Fuel_Spread0() * Fuel_PhiWind Depende del modelo de combustible Volver Depende del individuo

  23. Pendiente • Beta = acumulación carga/densidad de todas las partículas del modelo. • SlopeK = 5.275 * beta-0.3 • Fuel_PhiSlope = SlopeK * Slope2 • Rs = Fuel_Spread0() * Fuel_PhiSlope Depende del modelo de combustible Depende del individuo Volver

  24. CrossPoint mutación Algoritmo Genético • Algoritmo inspirado en la selección natural y en la genética. • - Trabaja sobre una población de individuos. • De forma iterativa, se evoluciona la población mediante las operaciones de: • Selección: competición de los individuos candidatos: los mejores individuos tienen mayor probabilidad de generar nuevos individuos. Función de evaluación. Elitismo. • Crossover: bajo una probabilidad se elije un crosspoint y los hijos reciben una parte de cada padre. • Mutación: bajo una probabilidad se muta el valor de un “cromosoma”. Población Nueva Población Crossover: Elitismo = 2 hijo1 Padre 1 hijo2 Padre 2 Volver

  25. - Init ∩ - Init Fitness = (∩ ∩) - Init) - Init) - Error = Real –Init Fitness - Error Volver

More Related