Improved Intelligent Optimization Algorithms And Research On Applications In Mid-and-Long Term Runoff Forecast

The result of runoff forecast is an important basis for water resources allocation,reservoir operation and flood prevention.With the long-term development of machine learning,algorithms such as support vector machine,which simulate human experience learning,are widely used in runoff forecasting to improve forecasting accuracy.Least squares support vector machine(LSSVM)is a modified algorithm based on the original support vector machine.It constructs an objective function based on the idea of least squares to reduce model parameters while speeding up the model learning process by solving equations.However,the performance of LSSVM is highly dependent on the choice of user-defined regularization parameter and kernel function parameter.Traditional parameter tuning methods are difficult to take into account both global optimization and local exploitation.Intelligent optimization algorithms begin to provide new idea for model tuning with strong robustness,and a runoff forecast model is established based on intelligent optimization algorithms to optimize LSSVM.However,the no free lunch theorem shows that it is impossible to solve all problems relying on only single optimization algorithm.Thus,the paper starts from two different types of optimization algorithms—Firefly algorithm(FA)and Yin-Yang-pair algorithm(YYPO),takes medium and long-term runoff forecast as the background,conducts different strategies on FA and YYPO to improve their comprehensive performance.Based on LSSVM model,combining with wavelet packet decomposition(WPD)to decompose and reconstruct hydrological data,the improved intelligent optimization algorithms are applied to tackle the real case of forecast.The main research contents and achievements are as follows.?Yin-Yang Firefly Algorithm(YYFA)is proposed.The opposition-based learning technique is used to initialize the firefly population to extract more information from the search space.Secondly,a greedy and full attraction model is adopted for the movement behavior of the firefly population.Finally,for the current optimal firefly after the movement,a single firefly is randomly generated in space to perform successive one-dimensional Cauchy mutation to deepen the capability of exploitation and exploration for the algorithm.The simulation results of 13 classic test functions show that YYFA performs in advance of other modified firefly algorithms.?Orthogonal Opposition-based-learning Yin-Yang-pair Algorithm(OOYO)is proposed.Retain the stages of splitting and archive in YYPO for local exploitation and eliminate redundant steps in YYPO that use weak point for global exploration.And an orthogonal opposition-based-learning strategy is designed,which is employed by the candidate point set generated by the preferable point focuses on exploitation to select the better from the better.Meanwhile,the initial radius and the random distribution in the splitting process is reanalyzed to complete the modification process.The simulation results of the CEC2020 test set show that the overall performance of OOYO is better than YYPO and other advanced algorithms.?WPD-YYFA-LSSVM and WPD-YYFA-LSSVM medium and long-term runoff forecast models are proposed.The annual runoff of the Yilihe river and Hulanhe river,and the monthly runoff of the Taolaihe River and Luohe river are used as the annual and monthly scale forecast research objects,respectively.The WPD technology is used to decompose each hydrological and meteorological element sequence,and the applicability of LASSO regression method is discussed based on the full factor input comparison.Based on the sub-sequence,the WPD-YYFA-LSSVM and WPD-OOYO-LSSVM model forecasts are superimposed and reconstructed to form a final runoff forecast.At the same time,other contrast models were established from the perspective of intelligent optimization algorithm and basic model respectively.The average evaluation results of several independent runs show that YYFA and OOYO perform well in the optimization of LSSVM model parameters,among which WPD-YYFA-LSSVM ranks at the top of the list.In the annual runoff forecast,the RMSEs of Yilihe river in training period and testing period are 26 million m~3 and 96 million m~3 respectively,and the RMSEs of Hulanhe river in training period and testing period are 0.15mm and 2.14mm respectively.As for the monthly scale,the RMSEs of training period and testing period in Taolaihe river is 1.05m~3/s and 0.88m~3/s respectively,and the RMSEs of training period and testing period about Luohe river is 5.39m~3/s and 3.76m~3/s respectively.The medium-and-long-term runoff forecasting models proposed in this paper can provide new ways for practical forecasting.