Napovedovanje cen zlitin s strojnim učenjem

magistrsko delo

Žiga Štosir (Author), Ljupčo Todorovski (Mentor)

Abstract

V magistrski nalogi smo razvili modele za kratkoročno napovedovanje cen fero zlitin oz. legur, ki se v jeklarski industriji uporabljajo za izdelavo visoko legiranega jekla. Za razvoj napovednih modelov smo uporabili predvsem ansambelske metode, kot je metoda naključnih gozdov (angl. random forests), v kombinaciji z analizo avtokorelacijskih funkcij časovnih vrst. Za primerjavo smo uporabili tudi metodo rekurenčnih nevronskih mrež (angl. recurrent neural network). Uspešnost modelov smo vrednotili s primerjavo z naivnim modelom nespremenjene vrednosti (angl. no-change model). Dodatno smo izvedli še optimizacijo hiperparametrov, s katero smo želeli uspešnost modelov dodatno izboljšati. Ugotovili smo, da se napovedni modeli ne odrežejo bistveno bolje od naivnega modela nespremenjene vrednosti. Možen razlog je, da naši napovedni modeli ne vključujejo dodatnih spremenljivk oz. dejavnikov kot so cene drugih surovin, svetovna proizvodnja, zaloge visoko legiranega jekla ter cene energentov, ki so pomembni za proizvodnjo legiranega jekla. V primeru, da bi želeli napovedni model uporabiti v praksi, bi ga bilo potrebno nadgraditi z upoštevanjem več spremenljivk.

Keywords

časovne vrste;napovedni modeli;strojno učenje;nakjlučni gozdovi;rekurenčne nevronske mreže;

Data

Language:	Slovenian
Year of publishing:	2024
Typology:	2.09 - Master's Thesis
Organization:	UL FRI - Faculty of Computer and Information Science
Publisher:	[Ž. Štosir]
UDC:	004.42
COBISS:	234924035
Views:	27
Downloads:	6
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Predicting alloy prices using machine learning
Secondary abstract:	In the master's thesis, we developed models for short-term forecasting of ferroalloy prices, which are used in the steel industry for the production of high-alloy steel. For the development of forecasting models, we employed ensemble methods, such as the random forests method, in combination with the analysis of autocorrelation functions of time series. For comparison, we also utilized the method of recurrent neural networks. The performance of the models was evaluated by comparing them to a baseline model that assumes no changes in price. Additionally, we performed hyperparameter optimization to further improve the models' performance. Our findings revealed that the forecasting models did not perform significantly better than the baseline model that assumes no price changes. One of the reasons is that the developed models do not incorporate additional variables or factors, such as the prices of other raw materials, global production levels, inventories of high-alloy steel, and energy prices, which are important for alloy steel production. If the forecasting model were to be used in practice, it would need to be enhanced by including additional variables.
Secondary keywords:	time series;predictive models;machine learning;random forests;recurrent neural network;
Type (COBISS):	Master's thesis/paper
Study programme:	0
Embargo end date (OpenAIRE):	1970-01-01
Thesis comment:	Univ. v Ljubljani, Fak. za matematiko in fiziko, Oddelek za matematiko, Računalništvo in matematika - 2. stopnja
Pages:	IX, 39 str.
ID:	26366411