Is there a way to constrain the estimated recharge (setup an upper bound)? I get an estimated recharge that is up to 90% of precipitation for the area that is expected to have a low recharge (less than 5 %).

[petertsaichiahsing]

Hi all,

Is there a way to setup a upper bound for the estimated recharge or constrain the recharge? I get an estimated recharge that is up to 90% of precipitation for the area that is expected to have a low recharge (less than 5 %).

I used daily precipitation, ET, and groundwater elevation data. I have used the Flexmodel as my recharge model.

Partial of codes are attached here.

ml = ps.Model(df_data,name='Trane')
noise_model = ps.ArNoiseModel()
ml.add_noisemodel(noise_model)
rch = ps.rch.FlexModel()

rm = ps.RechargeModel(prec_series, evap_series, recharge=rch, rfunc=ps.Gamma(), name="rch")

ml.add_stressmodel(rm)
ml.set_parameter("noise_alpha", vary=False, initial=0.001)
print(ml.parameters.loc['noise_alpha', 'initial'])

ml.solve(freq='D',tmin="5/15/2024", tmax="9/5/2024", solver=ps.LeastSquares(), report=False)

I also attached a screen shot of my fitting results here.

Looking forward to the discussion/answer here. Thank you, all.

Best regards,
Peter

[mbakker7]

You have a very short series (3 months) and get an amazing fit. But you are unhappy with the physical reality. It is good that you check that. The parameters of the FlexModel are inherently correlated. It is not clear from your output whether any of the parameters are fixed (I think lp and simax?). But if you, for example, fix srmax to 50 or 150, then kv will change but you will likely give a very similar fit. But both will give a different recharge.

My concern with the short series is that the response time is 100 days, which is about the length of your measurements. That will make it difficult to calibrate the model. Do you at least have precipitation and evap for a period before the first head measurement?

Last comment: the ArNoiseModel doesn't do anything (noise and residuals are the same and alpha is very small, probably lower bound), so you can remove the noise model (simpler model is better).