Energy usage optimization algorithms for cost-effective and resource-saving production are already widespread in industry. However, many optimization models disregard the uncertainty in the prediction of electricity and heat demand or in the forecast of solar energy feed-in. This paper presents a stochastic optimization model for a cross-sectoral energy system with multiple sources of uncertainty, combined with a new approach to generate scenarios for the optimization. The model maps the fluctuating electricity and heat consumption as well as the solar feed-in of a manufacturing plant with respect to the underlying uncertainty, which can be non-negligible. First, the focus will be on the generation of a set of equally likely scenarios displaying the uncertainty of the given predictions. In this case study, the electricity self-generation of a PV plant as well as the energy and heat demand by the production are classified as uncertain and considered in multiple scenarios. To create adequate scenarios, the new approach uses the probability density function (PDF), which will be calculated by a kernel density estimator (KDE) for each uncertain decision factor, based on the historical error between measurement and prediction. Afterwards, the selection of random samples of the error distribution is manipulated by a normal distribution to avoid large deviation between consecutive selected representatives. This leads to a set of scenarios in which implausible fluctuation is avoided. Thus, scenarios over- or underestimate the prediction or have a smooth transition between over- and underestimation. Based on the calculated scenarios, a stochastic optimization model was built and solved with respect to the modeled uncertainties. The extensive form is used to optimize as cost-effectively as possible, with the highest possible self-consumption of energy. The results are compared with the deterministic optimization model and evaluated based on the total costs and the necessary computational time. To further evaluate the new approach, state of the art sampling methods are used as benchmark.