- Comparison between model results and calibration data Looking back on the parameter calibration process, I tried to analyze the change trend of the simulation results through the image difference between the simulated value and the measured value of the model. At first, I wanted to simulate the data of one year, which would speed up the rhythm of trial and error. However, after a field test, it was found that the measured values of some indicators in the previous year and the following year fluctuated greatly and were unstable, so the two-year number should be used According to the comparison. In the process of parameter calibration, MP1 was modulated manually, and Chla was adjusted after MP1 was stable. The distribution of Chla at different times was plotted. （1） Chla
Firstly, the comparison results between the measured data and the simulated values of Chla are presented. In order to accurately grasp the differences of each station, different results of s1-s20 are drawn, as shown in the figure below. The blue scatter points are the measured data, and the red discount is the simulation value. It can be seen from the figure below that the simulation effect of algae is better.
Secondly, in order to determine which algae growth results will affect the judgment, and accurately find out the parameters affecting the growth of this algae in the model and change the final model results, I choose to further draw the subdivision drawing of cyanobacteria, green algae and diatoms. The yellow block indicates that the corresponding block color of diatoms and cyanobacteria is blue, and that of green algae is green. As can be seen from the figure below, in my model results, diatoms grow crazily in the pre sequence time, while cyanobacteria grow in the later period, and the proportion of green algae is not large.
（2） MP1
Aquatic vegetation is very important, it significantly affects the growth of algae, and will change the nutrients in the water and its change trend, so the first parameter I adjust is MP1. The process of MP1 adjustment is a trial and error process. I have drawn a total of 800 MP1 images. Finally, the comparison of MP1 measured and simulated in the optimal results is as follows:
（3） NH4
The comparison of NH4 measured and simulated in the optimized results is as follows:
（4） No3
The comparison of NO3 measured and simulated in the optimized results is as follows:
（5） PO4
The comparison of PO4 measured and simulated in the optimized results is as follows:
（6） TN
The comparison of TN measured and simulated in the optimized results is as follows:
（7） TP
The comparison of TP measured and simulated in the optimized results is as follows:
2. Algorithm idea, algorithm construction, implementation process, convergence process and algorithm efficiency
（1） Algorithm ideas
My result is the combination of manual calibration and automatic calibration.
Automatic calibration can not be used completely, because some key links of the model need to be understood manually step by step, and the automatic calibration may lead to the same effect of different parameters. Even if the very unreasonable parameter range may have a positive impact on the results, it is unnecessary and helpless to help us understand the model as a whole. Therefore, I first manually adjusted the model parameters of aquatic vegetation and algae, which greatly improved the initial setting of the model. My initial default Nash coefficient was about - 3.
I first adjusted TMC1, tmc2, tmg1, tmg2, tmd1, tmd2, bmrg, BMRC, bmrd, PMG, PMC, PMD, WSD, WSG, WSC, prrc, PRRD, prrg, khnc, khnd, khng, khpc, khpd, khpg, IWD_ KHNm、IWD_ KHPm、IWD_ KHNs、IWD_ KHPs、IWD_ DOPTm、VEGH、VEGBM、VEGTM、KPHYTO、KBP、IWD_ PCR、IWD_ PMm、IWD_ BMRm、IWD_ PRRm、IWD_ SETM、IWD_ TMm1、IWD_ TMm2、IWD_ Ktg1m, which is related to aquatic vegetation or algae, manually revised the parameters of algae and MP1 repeatedly by observing the comparison chart between the simulated value and the measured value, and finally achieved high Nash coefficient, in which algae almost reached 0.6, MP1 was close to 0.7, but after adding nutrients, the coefficient of MP1 and algae decreased a little.
After that, kthdr, ktmnl, khdnn, andc, rnitm, khnitn, knit2, KDN, wsrp, wslp, rnpref, KDC, Krn, KLN, KRP, klp, KDP, KRC, KLC, AONT and other nutrient parameters were automatically adjusted by the algorithm.
Derivative is the optimization based on gradient. Only when the function form is known can the derivative be obtained, and the function can be convex function. However, in fact, these two conditions are not satisfied in many cases, so gradient optimization can not be used, especially for this scenario, because the calculation process through iwind-lr is a black box, and the calculation time consumption is too long to obtain accurate function form. Therefore, I choose to use kappa optimization and set its threshold value to 0.5, which shows that the optimization process is more biased To explore boldly.
（2） Algorithm construction
The black box
It's input data
Is the acquisition function
It is a model based on input data hypothesis, that is, the known input data is on this model. There are many models that can be used to assume, such as Gaussian model.
Initialize the dataset: this step is to initialize the fetched dataset
Cycle parameter selection times
Because every time the parameters are selected, it needs to be calculated. Every time a function is calculated, a large amount of resources will be consumed. Therefore, it is generally necessary to fix the number of parameter selection (or function evaluation times).
Computing Gaussian processes
Select sampling function (UCB)
Entropy search
Upper bound of confidence
（3） Convergence process
Due to the time limit, the model converges to 836 times (the running time is 20 days), and the maximum comprehensive Nash coefficient is 0.2573147.
（4） Algorithm efficiency
In addition, there is a set of mechanism to save and automatically contact the pre sequence running, which is very effective