#!/usr/bin/env python
# coding: utf-8

# #  Time interpolation

# In[1]:


import numpy as np
import mikeio


# In[2]:


ds = mikeio.read("../tests/testdata/waves.dfs2")
ds


# ## Interpolate to specific timestep
# 
# A common use case is to interpolate to a shorter timestep, in this case 1h.

# In[3]:


ds_h = ds.interp_time(3600)
ds_h


# And to store the interpolated data in a new file.

# In[4]:


ds_h.to_dfs("waves_3h.dfs2")


# ## Interpolate to time axis of another dataset

# Read some non-equidistant data typically found in observed data.

# In[5]:


ts = mikeio.read("../tests/testdata/waves.dfs0")
ts


# The observed timeseries is longer than the modelled data. Default is to fill values with NaN.

# In[6]:


dsi = ds.interp_time(ts)


# In[15]:


dsi.time


# In[8]:


dsi["Sign. Wave Height"].shape


# In[9]:


ax = dsi["Sign. Wave Height"].sel(x=250, y=1200).plot(marker='+')
ts["Sign. Wave Height"].plot(ax=ax,marker='+')


# ## Model validation
# 
# A common metric for model validation is mean absolute error (MAE).
# 
# In the example below we calculate this metric using the model data interpolated to the observed times.
# 
# For a more elaborate model validation library which takes care of these things for you as well as calculating a number of relevant metrics, take a look at [fmskill](https://github.com/DHI/fmskill).
# 
# Use `np.nanmean` to skip NaN.

# In[10]:


ts["Sign. Wave Height"]


# In[11]:


dsi["Sign. Wave Height"].sel(x=250, y=1200)


# In[12]:


diff = (ts["Sign. Wave Height"]  - dsi["Sign. Wave Height"].sel(x=250, y=1200))
diff.plot()


# In[13]:


mae = np.abs(diff).nanmean().to_numpy()
mae


# # Clean up

# In[14]:


import os
os.remove("waves_3h.dfs2")