I have produced calibration (1984-1988 )and validation (1989-1993) results, but their results are inverted. I used the parameters that was derived from the calibration model and applied to the validation model. So, how can I optimize my result?
What do you mean by "their results are inverted"? Assuming that you mean that the validation period gives a better result than the calibration period, this is a little unusual, but not impossible. For hydrological models, there could be a significant difference in the drivers (i.e. weather) between the calibration and validation periods. For example, if the calibration period included both floods and droughts, and the validation period was more normal, then it is quite possible that the validation statistics would be better than the calibration statistics.
The RB and NSE of my calibration period is -4.16 and 0.69, respectively. After I applied the same parameters value to validation period, the RB and NSE become 20.1 and 0.3, respectively.
I assume that RB is the relative bias - probably reported as a percentage? This means there is a significant bias in the validation period, which probably explains the lower NSE. You have selected a 5 year period for the calibration, and have a 5 year period for validation.
Can you give me information on the catchment you are modelling (e.g. climate zone, area, dominant land use) and which model you are using? Also, is there any significant difference in the weather patterns between the calibration and validation periods?
Yes, RB is relative bias expressed in percentage. My study area in tropical region (Malaysia), latitudes 4° to 6° and longitudes 101° to 103°. the annual precipitation is 2505 mm. The northeast monsoon (NEM) season brings heavy precipitation to the basin from November to January. While, the basin receives lesser precipitation during the southwest monsoon (SWM) season between May and August. three main land uses/land covers in KRB: forest (76%), oil palm (11%) and rubber (11%). The model that I used is Rainfall-Runoff-Inundation (RRI) Model. during the calibration and validation period, there is no significant difference in weather patterns.
I am not that familiar with the RRI model, so can't talk from experience there. You said there was no significant difference in the weather patterns between the two periods - how variable is the annual total precipitation between years? Are there any years in the validation period that has significant more or less rainfall than in the calibration period? Also, are you evaluating the model based on just the runoff, or considering inundation as well?
during calibration period, annual total precipitation is 2365 mm. validation period, annual total precipitation is 2514 mm. So two periods have similar total precipitation. I just evaluate the model based on runoff due to lack of spatial data for evaluating inundation.
The average annual rainfall may be similar between the calibration and validation periods, but is there any single year with significantly higher or lower rainfall. In one sense, the tropics are an easier area to model as the general pattern doesn't vary much from year to year. However, given the impact of large single events including tropical cyclones (aka hurricanes or typhoons), there is the possibility of an extreme event which the model will not capture well. Further, estimates of the rainfall for such events is very problematic due to the strong winds affecting rain gauges. There is also a problem measuring extremely high flows. All these factors can lead to a poor model. The question is whether there is an extreme event somewhere in the validation period?
If you are only interested in modelling streamflow, then a simpler rainfall-runoff model may be more appropriate. There are quite a few to select from. My group produced the hydromad package (https://hydromad.catchment.org/). This is an R package that includes several rainfall-streamflow models, including the GR4J, IHACRES, SIMHYD ans Sacramento models.
So both the calibration and validation periods include a flood event - are they similar magnitudes? Do you run the model over both the calibration and validation periods and extract the stats for the validation period (this is better than two separate runs of the model).
Final question - how much of a warm up period are you using? This is a period of time at the start of the simulation which you don't use to assess the model. It allows the model to forget the initial values for storages. Fast decaying storages (e.g. quick flow) forget their initial value very quickly. The issue is the slowly decaying storages (e.g. slow flow or baseflow).
Yes, both events have similar magnitudes in term of total rainfall. I haven't run the model over both calibration and validation but will run for it. Is it acceptable, I just run my simulation over calibration and validation periods if I fail to obtain acceptable result for calibration and validation periods?
For the warm up period, I excluded the first six months in five years simulation in which simulated hydrograph has correlation to observed.
You should always strive to calibrate your model with the recent and enough data. So for me you should have used some of the data in 1984-1993 data for calibration and validation you could use some of data in 1984-1993. There is no such thing as validation. If you do right calibration no need for validation. You can use 2021 to validate your data.
I am not sure if the ratio between calibration and validation should not be different, e. g. 70:30 (cal:val). Longer calibration data set can describe the study area better to obtain more precise model parameters in optimisation process.
You should first look to If there is any single year with significantly higher or lower rainfall may be this year make the difference.
Usually, Model calibration means the estimation of those parameters from historical input-output records. Model validation means judging the performance of the calibrated model over that portion of historical records which have not been used for the calibration.
Why you divide your data as a calibration (1984-1988 ) and validation (1989-1993) to be almost half and and half?