3.1. Wolf Pack Algorithm (WPA)
The Wolf Pack Algorithm is a novel swarm intelligence algorithm with strong local search ability and global convergence. It consists of the leading wolf, fierce wolves, and explore wolves, including three kinds of intelligent behaviors: Scouting, summoning, and beleaguering. Simultaneously, with a productive rule for the leading wolf, which is that the winner can dominate all, it is a renewable mechanism, namely survival of the fittest, for a pack of wolves. It has been applied in the field of mathematics, physics, and hydropower station optimization, achieving good calculation results [
29,
30,
31]. In this paper, the WPA is used for the parameter estimation of the Muskingum model. Combining the principle of the WPA with the objective function, we designed a WPA for solving Muskingum model parameters. The procedure of our algorithm for parameter optimal estimation of the Muskingum model is shown as follows:
Step 1: (Initial wolves) We must initially determine the size of the algorithm population
and the number of iterations
, the probe wolf scale factor
, the population regeneration factor
, the step factor
, the distance determination factor
, and the maximum number of scouting
. The Model Parameter
is regarded as the position of the artificial wolf
in a two-dimensional decision space. The position of the wolf pack is initialized in the range of
and
, as shown in Equation (7):
where
is the initial position of the
i-th wolf in the
j-th decision space;
is a uniformly distributed random number in the interval [0,1].
Step 2: (Scouting) Choose the objective function f as the prey odor concentration and calculate the prey odor concentration at the location of the artificial wolf. According to the , sort all artificial wolf positions in descending order, then select the first artificial wolf of the sorted population as a head wolf, whose location is regarded as and the prey odor concentration is . Choose the second to artificial wolves as the explore wolves that total , and the remaining artificial wolves as the fierce wolves. The explore wolves scout to conduct a fine search, until the maximum number of walks is reached or the maximum global optimal solution is found.
Step 3: (Summoning) The fierce wolfs quickly approach the head wolf to achieve the global convergence of solution, until the Euclidean distance
between all fierce wolves and the head wolf is less than the judgment value of distance
or a contemporary global optimal solution is found. The calculation formula of the distance judgment value is:
where
is the judgment value of distance;
is the spatial dimension;
is the distance determining factor;
,
are the upper and lower limits of the decision variable
, respectively.
Step 4: (Sieging) Under the command of the head wolf, other artificial wolves further update the location to achieve local fine optimization.
Step 5: (Competitive update) According to the population, update factor b, calculate the amount of the population, update and regenerate artificial wolves, which can be alternated with the previous generation of artificial wolves with a low fitness value to ensure the diversity of solutions and to avoid falling into the local optimal.
Step 6: Determine whether the maximum number of iterations has been reached; if it has, output the optimal global solution. If not, proceed to the next generation calculation and return to Step 2 until the maximum number of iterations is reached.
The WPA flow chart for solving the Muskingum model parameters is shown in
Figure 2.
3.2. Parameter Sensitivity Analysis
The WPA involves a relatively large number of parameters; the main sensitivity parameters are the distance determination factor and the step factor . Therefore, this paper focuses on the discussion of parameters and on the algorithm. In order to determine the distance determination factor and the step factor , the parameters were set as follow: Artificial wolf population size , explore wolf proportion factor , population update factor , the maximum number of scouting , the number of iterations . With the Muskingum principle, the algorithm is used to solve -dimensional space problem , and the upper and lower limits of the variables are and , respectively.
The WPA’s raid step:
where
is the step size of the
d-dimensional space;
is the step factor;
,
are the upper and lower limits of the decision variable
, respectively.
In the summoning act, the condition of the artificial wolf raid termination is
, so the rapid step should meet the following formula:
where the parameters are the same as above.
From Formula (10):
where the parameters are the same as above.
From the above derivation, it can be seen that the normal operation of the algorithm can be satisfied when the step factor and the distance determination factor satisfy the formula (11).
Fixed distance determination factor
, set
, and 130, respectively. For each
, run the program code 20 times independently, select the evaluation index of the algorithm, the absolute deviation of the flood process, and the observed and simulated flood peak deviation. The absolute deviation of the flood process and the formula are as follow:
where
is the absolute deviation of the flood process, m
3/s.
The flood peak deviation is:
where
f is the flood peak deviation;
and
are the observed and simulated maximum outflow at peak flow event number
i, respectively;
n is the number of simulations times.
The evaluation index values are taken from the average of 20 independent runs. Taking the flood event of 20100821 as an example, the results are shown in
Table 3.
(1) With the increase of the value, decreases slightly from 7892.7 m3/s to 7891.6 m3/s, which indicates that the larger the step size, the finer the search, and the closer the flood forecast results are to the actual flood process.
(2) When , the variation range of and flood peak are not large, which indicates that the step size of the WPA has little influence on the forecast results in a certain range.
Keeping in mind that the smaller the searching step, the more time consuming the process is, the initial selection of
, that is
.
, and 70 are selected to simulate the same flood, and the results are shown in
Table 4.
(1) With the increase of , decreases slightly, and the minimum value appears at . The peak deviation is also minimized, and then the value increases; the optimization effect has no obvious change.
(2) The range of is not significant, and indicates that the parameter of the wolves algorithm on the two-dimensional nonlinear optimization problem has little influence on the algorithm, and the algorithm is more applicable.
With the increase of , the raid step will be smaller, and the optimization result is too fine, which will cause the artificial wolf to turn from difficult to besieging. Thus, the algorithm has the possibility of entering an infinite loop.
In order to further analyze the influence of different weight coefficients on the forecast results, we set different weight coefficients using the same method. The results are shown in
Table 5.
(1) With the decrease of , peak deviation increased slightly, with an increase 0.1 m3/s only. When is 0.8 or 0.9, the flood peak deviation takes the minimum, which indicated that the weight factor of Objective 1 focuses on the minimum deviation of flood peak.
(2) With the increase of , the absolute deviation of the flood process is gradually reduced from 7916.9 m3/s to 7891.6 m3/s, and reaches the minimum value when is 0.7, which indicated that the weight factor of Objective 2 focuses on the total error of the entire flood process.
According to comprehensive analysis, the weight coefficient had a great influence on the fitting effect of the whole flood process and had little effect on the flood peak simulation. Therefore, and were selected for the multi-objective weight value.