With the increasing share of variable renewable energy sources in the power system, electricity prices are becoming more and more volatile and uncertain. This means that electricity market participants are experiencing issues related to trading activities as wrong electricity forecast can lead to wrong schedules and reduced profits. The state-of-the-art literature offers wide range of electricity price forecasting tools which build upon historical prices as well as various exogenous variables as inputs. The deep neural network algorithms are prevailing in literature and impose as the most advanced tools. In this paper we propose a long-short-term-memory network using only historic price, its timestamp and additional features engineered on top of that price. We will elaborate which features affect the forecasting and show how the algorithms trained on prices before significant change in trends behave when unexpected prices occur, such as those in the second half of 2021. The first research outcome is that proper feature selection can have significant impact on forecasting accuracy, for example day-of-week and statistical properties of last 24 or 168 hours increase the accuracy noticeably. The second one is that when drastic trend changes occur in the historical data it may show that using last few months as a test dataset is not a proper way to handle the issue. Better results are achieved if the test dataset is taken as certain months during the year.