fmda/moisture_models.py

   1 import numpy as np
   2 import math
   3 import matplotlib.pyplot as plt
   4 import copy
   5 from abc import ABC, abstractmethod
   6 import xgboost as xg
   7 from xgboost import XGBRegressor
   8 from sklearn.metrics import mean_squared_error
   9 import pandas as pd
  10 from sklearn.ensemble import RandomForestRegressor
  11 from sklearn.linear_model import LinearRegression
  12 from utils import Dict
  13
  14 # ODE + Augmented Kalman Filter Code
  15 #~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  16
  17 def model_decay(m0,E,partials=0,T1=0.1,tlen=1):
  18     # Arguments:
  19     #   m0          fuel moisture content at start dimensionless, unit (1)
  20     #   E           fuel moisture eqilibrium (1)
  21     #   partials=0: return m1 = fuel moisture contents after time tlen (1)
  22     #           =1: return m1, dm0/dm0
  23     #           =2: return m1, dm1/dm0, dm1/dE
  24     #           =3: return m1, dm1/dm0, dm1/dE dm1/dT1
  25     #   T1          1/T, where T is the time constant approaching the equilibrium
  26     #               default 0.1/hour
  27     #   tlen        the time interval length, default 1 hour
  28
  29     exp_t = np.exp(-tlen*T1)                  # compute this subexpression only once
  30     m1 = E + (m0 - E)*exp_t                   # the solution at end
  31     if partials==0:
  32         return m1
  33     dm1_dm0 = exp_t
  34     if partials==1:
  35         return m1, dm1_dm0          # return value and Jacobian
  36     dm1_dE = 1 - exp_t
  37     if partials==2:
  38         return m1, dm1_dm0, dm1_dE
  39     dm1_dT1 = -(m0 - E)*tlen*exp_t            # partial derivative dm1 / dT1
  40     if partials==3:
  41         return m1, dm1_dm0, dm1_dE, dm1_dT1       # return value and all partial derivatives wrt m1 and parameters
  42     raise('Bad arg partials')
  43
  44
  45 def ext_kf(u,P,F,Q=0,d=None,H=None,R=None):
  46     """
  47     One step of the extended Kalman filter.
  48     If there is no data, only advance in time.
  49     :param u:   the state vector, shape n
  50     :param P:   the state covariance, shape (n,n)
  51     :param F:   the model function, args vector u, returns F(u) and Jacobian J(u)
  52     :param Q:   the process model noise covariance, shape (n,n)
  53     :param d:   data vector, shape (m). If none, only advance in time
  54     :param H:   observation matrix, shape (m,n)
  55     :param R:   data error covariance, shape (n,n)
  56     :return ua: the analysis state vector, shape (n)
  57     :return Pa: the analysis covariance matrix, shape (n,n)
  58     """
  59     def d2(a):
  60         return np.atleast_2d(a) # convert to at least 2d array
  61
  62     def d1(a):
  63         return np.atleast_1d(a) # convert to at least 1d array
  64
  65     # forecast
  66     uf, J  = F(u)          # advance the model state in time and get the Jacobian
  67     uf = d1(uf)            # if scalar, make state a 1D array
  68     J = d2(J)              # if scalar, make jacobian a 2D array
  69     P = d2(P)              # if scalar, make Jacobian as 2D array
  70     Pf  = d2(J.T @ P) @ J + Q  # advance the state covariance Pf = J' * P * J + Q
  71     # analysis
  72     if d is None or not d.size :  # no data, no analysis
  73         return uf, Pf
  74     # K = P H' * inverse(H * P * H' + R) = (inverse(H * P * H' + R)*(H P))'
  75     H = d2(H)
  76     HP  = d2(H @ P)            # precompute a part used twice
  77     K   = d2(np.linalg.solve( d2(HP @ H.T) + R, HP)).T  # Kalman gain
  78     # print('H',H)
  79     # print('K',K)
  80     res = d1(H @ d1(uf) - d)          # res = H*uf - d
  81     ua = uf - K @ res # analysis mean uf - K*res
  82     Pa = Pf - K @ d2(H @ P)        # analysis covariance
  83     return ua, d2(Pa)
  84
  85 ### Define model function with drying, wetting, and rain equilibria
  86
  87 # Parameters
  88 r0 = 0.05                                   # threshold rainfall [mm/h]
  89 rs = 8.0                                    # saturation rain intensity [mm/h]
  90 Tr = 14.0                                   # time constant for rain wetting model [h]
  91 S = 250                                     # saturation intensity [dimensionless]
  92 T = 10.0                                    # time constant for wetting/drying
  93
  94 #~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  95
  96 def model_moisture(m0,Eqd,Eqw,r,t=None,partials=0,T=10.0,tlen=1.0):
  97     # arguments:
  98     # m0         starting fuel moistureb (%s
  99     # Eqd        drying equilibrium      (%)
 100     # Eqw        wetting equilibrium     (%)
 101     # r          rain intensity          (mm/h)
 102     # t          time
 103     # partials = 0, 1, 2
 104     # returns: same as model_decay
 105     #   if partials==0: m1 = fuel moisture contents after time 1 hour
 106     #              ==1: m1, dm1/dm0
 107     #              ==2: m1, dm1/dm0, dm1/dE
 108
 109     if r > r0:
 110         # print('raining')
 111         E = S
 112         T1 =  (1.0 - np.exp(- (r - r0) / rs)) / Tr
 113     elif m0 <= Eqw:
 114         # print('wetting')
 115         E=Eqw
 116         T1 = 1.0/T
 117     elif m0 >= Eqd:
 118         # print('drying')
 119         E=Eqd
 120         T1 = 1.0/T
 121     else: # no change'
 122         E = m0
 123         T1=0.0
 124     exp_t = np.exp(-tlen*T1)
 125     m1 = E + (m0 - E)*exp_t
 126     dm1_dm0 = exp_t
 127     dm1_dE = 1 - exp_t
 128     #if t>=933 and t < 940:
 129     #  print('t,Eqw,Eqd,r,T1,E,m0,m1,dm1_dm0,dm1_dE',
 130     #        t,Eqw,Eqd,r,T1,E,m0,m1,dm1_dm0,dm1_dE)
 131     if partials==0:
 132         return m1
 133     if partials==1:
 134         return m1, dm1_dm0
 135     if partials==2:
 136         return m1, dm1_dm0, dm1_dE
 137     raise('bad partials')
 138
 139 def model_augmented(u0,Ed,Ew,r,t):
 140     # state u is the vector [m,dE] with dE correction to equilibria Ed and Ew at t
 141     #
 142     m0, Ec = u0  # decompose state u0
 143     # reuse model_moisture(m0,Eqd,Eqw,r,partials=0):
 144     # arguments:
 145     # m0         starting fuel moistureb (1)
 146     # Ed         drying equilibrium      (1)
 147     # Ew         wetting equilibrium     (1)
 148     # r          rain intensity          (mm/h)
 149     # partials = 0, 1, 2
 150     # returns: same as model_decay
 151     #   if partials==0: m1 = fuel moisture contents after time 1 hour
 152     #              ==1: m1, dm0/dm0
 153     #              ==2: m1, dm1/dm0, dm1/dE
 154     m1, dm1_dm0, dm1_dE  = model_moisture(m0,Ed + Ec, Ew + Ec, r, t, partials=2)
 155     u1 = np.array([m1,Ec])   # dE is just copied
 156     J =  np.array([[dm1_dm0, dm1_dE],
 157                    [0.     ,     1.]])
 158     return u1, J
 159
 160
 161 ### Default Uncertainty Matrices
 162 Q = np.array([[1e-3, 0.],
 163             [0,  1e-3]]) # process noise covariance
 164 H = np.array([[1., 0.]])  # first component observed
 165 R = np.array([1e-3]) # data variance
 166
 167 def run_augmented_kf(dat0,h2=None,hours=None, H=H, Q=Q, R=R):
 168     dat = copy.deepcopy(dat0)
 169
 170     if h2 is None:
 171         h2 = int(dat['h2'])
 172     if hours is None:
 173         hours = int(dat['hours'])
 174
 175     d = dat['y']
 176     feats = dat['features_list']
 177     Ed = dat['X'][:,feats.index('Ed')]
 178     Ew = dat['X'][:,feats.index('Ew')]
 179     rain = dat['X'][:,feats.index('rain')]
 180
 181     u = np.zeros((2,hours))
 182     u[:,0]=[0.1,0.0]       # initialize,background state
 183     P = np.zeros((2,2,hours))
 184     P[:,:,0] = np.array([[1e-3, 0.],
 185                       [0.,  1e-3]]) # background state covariance
 186     # Q = np.array([[1e-3, 0.],
 187     #             [0,  1e-3]]) # process noise covariance
 188     # H = np.array([[1., 0.]])  # first component observed
 189     # R = np.array([1e-3]) # data variance
 190
 191     for t in range(1,h2):
 192       # use lambda construction to pass additional arguments to the model
 193         u[:,t],P[:,:,t] = ext_kf(u[:,t-1],P[:,:,t-1],
 194                                   lambda uu: model_augmented(uu,Ed[t],Ew[t],rain[t],t),
 195                                   Q,d[t],H=H,R=R)
 196       # print('time',t,'data',d[t],'filtered',u[0,t],'Ec',u[1,t])
 197     for t in range(h2,hours):
 198         u[:,t],P[:,:,t] = ext_kf(u[:,t-1],P[:,:,t-1],
 199                                   lambda uu: model_augmented(uu,Ed[t],Ew[t],rain[t],t),
 200                                   Q*0.0)
 201       # print('time',t,'data',d[t],'forecast',u[0,t],'Ec',u[1,t])
 202     return u
 203
 204 # General Machine Learning Models
 205 #~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 206
 207 class MLModel(ABC):
 208     def __init__(self, params: dict):
 209         self.params = Dict(params)
 210         if type(self) is MLModel:
 211             raise TypeError("MLModel is an abstract class and cannot be instantiated directly")
 212         super().__init__()
 213
 214     def filter_params(self, model_cls):
 215         """Filters out parameters that are not part of the model constructor."""
 216         model_params = self.params.copy()
 217         valid_keys = model_cls.__init__.__code__.co_varnames
 218         filtered_params = {k: v for k, v in model_params.items() if k in valid_keys}
 219         return filtered_params
 220
 221
 222     def fit(self, X_train, y_train, weights=None):
 223         print(f"Fitting {self.params.mod_type} with params {self.params}")
 224         self.model.fit(X_train, y_train, sample_weight=weights)
 225
 226     def predict(self, X):
 227         print(f"Predicting with {self.params.mod_type}")
 228         preds = self.model.predict(X)
 229         return preds
 230
 231     def eval(self, X_test, y_test):
 232         preds = self.predict(X_test)
 233         rmse = np.sqrt(mean_squared_error(y_test, preds))
 234         # rmse_ros = np.sqrt(mean_squared_error(ros_3wind(y_test), ros_3wind(preds)))
 235         print(f"Test RMSE: {rmse}")
 236         # print(f"Test RMSE (ROS): {rmse_ros}")
 237         return rmse
 238
 239 class XGB(MLModel):
 240     def __init__(self, params: dict):
 241         super().__init__(params)
 242         model_params = self.filter_params(XGBRegressor)
 243         self.model = XGBRegressor(**model_params)
 244         self.params['mod_type'] = "XGBoost"
 245
 246     def predict(self, X):
 247         print("Predicting with XGB")
 248         preds = self.model.predict(X)
 249         return preds
 250
 251 class RF(MLModel):
 252     def __init__(self, params: dict):
 253         super().__init__(params)
 254         model_params = self.filter_params(RandomForestRegressor)
 255         self.model = RandomForestRegressor(**model_params)
 256         self.params['mod_type'] = "RandomForest"
 257
 258 class LM(MLModel):
 259     def __init__(self, params: dict):
 260         super().__init__(params)
 261         model_params = self.filter_params(LinearRegression)
 262         self.model = LinearRegression(**model_params)
 263         self.params['mod_type'] = "LinearRegression"
 264
 265
 266
 267
 268
 269
 270
 271
 272
 273
 274
 275
 276
 277
 278
 279
 280
 281
 282