Krzysztof Siwek

Photovoltaic forecasting studies often assume access to plant metadata and numerical weather predictions. A downstream energy manager may instead receive only an issued production trajectory and recent measurements. This study tests whether this trajectory can be corrected without coordinates, capacity, orientation, or forecast weather. The EMSx data provide energy measured every 15 minutes and vendor forecasts for 70 anonymised systems. Complete daily profiles spanning 24 hours are formed, using chronological training, tuning, calibration, and test partitions. The test contains 8,898 profiles and 854,208 matched deliveries. A global LightGBM model predicts residuals from the vendor trajectory, calendar variables, causal measurements, and recent forecast error. It reduces pooled mean absolute error from 71.34 to 52.18 kWh and improves all 70 systems. Most of the gain appears when the complete issued profile is added. Causal rolling calibration improves the grid continuous ranked probability score from 38.62 to 35.85 kWh and raises central 90% coverage from 0.827 to 0.919, although the weighted interval score changes from 48.91 to 49.24 kWh. After limited EMSx adaptation, the strongest foundation model variant, Chronos-2 with the vendor trajectory, reaches test MAE 61.34±0.11 kWh across 32 seeds. The residual correction retains lower test MAE.

Automated media bias detection is becoming more important as disinformation spreads faster than human fact-checkers can respond. Yet most NLP research treats model accuracy as the only measure of success, ignoring training cost and energy use. This study asks a different question: is the accuracy gain from larger, more complex transformer models actually worth the environmental and computational price? We trained and evaluated three transformer architectures like RoBERTa-base, RoBERTa-large, and DeBERTa-v3-large on a three-class bias classification task using the News Bias Data corpus. Each model was tested across multiple hyperparameter configurations, including Focal Loss, class-balanced weights, and learning rate adjustments. A TF-IDF baseline provided a classical reference point. CO₂ emissions were estimated using the Green Algorithms calculator, based on the hardware specifications of the training environment. DeBERTa-v3-large with Focal Loss, class-balanced weights, and a reduced learning rate achieved the best classification result, with a Macro F₁ of 0.9756. The cost was steep: training took around 140 hours and produced more than four times the emissions of the Pareto-efficient RoBERTa-large setup. That setup was with Focal Loss and reached a Macro F₁ of 0.9487 for 9.95 kg CO₂e. The TF-IDF baseline used 99.9% less energy but fell 20–28 percentage points short on accuracy. The relationship between model quality and computational cost is strongly non-linear. Gains beyond a certain threshold require disproportionate resources. For most real-world applications, a well-tuned RoBERTa-large is a more responsible choice than the top-performing model available.

Precise prediction of photovoltaic (PV) energy generation is essential for op-timal, profitable and ecological management of electric energy resources all over the world. As a result, attempts are being made to develop more accurate prediction algo-rithms. This paper compares the application of Long Short-Term Memory (LSTM, a sub-type of Recurrent Neural Networks), PatchTST (a type of Transformer Neural Network - TNN) and ensemble models (making use of these two approaches) for estimating PV en-ergy production 24 hours ahead. The results indicate that both analysed single methods have comparable prediction accuracy, though the hybrid approach outperforms them. The experiments were conducted on data from PV sites deployed across campuses at Austral-ian La Trobe University. However, future studies could verify this approach using differ-ent datasets. Algorithms and results presented in this study may especially contribute to the development of Recurrent and Transformer Neural Networks as prediction methods of PV energy production.

The aim of the study was to determine the causes of eutrophication in small urban water reservoirs located in the UMCS Botanical Garden in Lublin, supplied via surface and groundwater. The research (hydrological, hydrobiological, and hydrochemical), which included both field and laboratory work, was conducted during the growing season in the years 2022–2023. These ponds are fed by waters from the Czechówka River and, to a lesser extent, by groundwater (seepage). Both river and groundwater are characterised by high concentrations of mineral forms of nitrogen and phosphorus, ranging from 1.49 to 12.0 mg N∙dm–3 and 0.07 to 0.21 mg P∙dm–3, respectively. This contributes to the intensive development of phytoplankton, especially during the summer period, with diatoms dominating the phytoplankton structure. The trophic state of the ponds ranges from eutrophic to hypertrophic. The study showed that the high degree of eutrophication was due to the load of nutrients delivered by the waters of the Czechówka River. Despite having several times higher concentrations of mineral and total nitrogen than in the river water, the spring water feeding the ponds had a negligible impact on the quality of the pond water due to low flow rate (<0.5 dm3∙s−1). Therefore, the construction of urban ponds as part of green-blue infrastructure should consider the possibility of reducing nutrients through biogeochemical barriers and suspended sedimentation. It is also advisable to partially shade the water surface by planting trees in the shoreline zone to limit water heating and phytoplankton development.

Magnetic Resonance Imaging is increasing in importance in prostate cancer diagnosis due to the high accuracy and quality of the examination procedure. However, this process requires a time-consuming analysis of the results. Currently, machine vision is widely used in many areas. It enables automation and support in radiological studies. Successful detection of primary prostate tumors depends on the effective segmentation of the prostate itself. At times, a CT scan may be performed; alternatively, MRI may be the selected option. The data always reach a bottleneck stage. This paper presents the effective training of deep learning models to segment the prostate based on onefold and multimodal medical images. This approach supports the computer-aided diagnosis (CAD) system for radiologists as the first step in cancer exams. A comparison of two approaches designed for prostate segmentation is described. The first combines YOLOv4, the object detection neural network, and U-Net for a semantic segmentation based on onefold modality MRI images. The second presents the same method trained on multimodal images—a CT and MRI mixed dataset. The learning process was carried out in a cloud environment using GPU cards. The experiments are based on data from 120 patients who have undergone MRI and CT examinations. Several metrics evaluated the trained models. In the prostate semantic segmentation process, better results were achieved by mixed MRI with CT datasets. The best model achieved the value of 0.9685 for the Sørensen–Dice coefficient for the threshold value of 0.6.

The shunt static compensator with the control based on the average signals is presented. Two control signals are considered - moving RMS value of grid voltage and moving average power of the energy source. The compensator generates reactive current computed in the real time with the use of the iterative method. The multiport representation of the grid is presented. Analytical and simulation results show how the iterative control of the compensator behaves for three kinds of control signals.

Background: The purpose of the paper is to propose different arrangements of neural networks for short-time 24-h load forecasting in Power Systems. Methods: The study discusses and compares different techniques of data processing, applying the feedforward and recurrent neural structures. They include such networks as multilayer perceptron, radial basis function, support vector machine, self-organizing Kohonen networks, deep autoencoder, and recurrent deep LSTM structures. The important point in getting high-quality results is the composition of many solutions in the common ensemble and their fusion to create the final forecast of time series. The paper considers and compares different methods of fusing the individual results into the final forecast, including the averaging, application of independent component analysis, dynamic integration, and wavelet transformation. Results: The numerical experiments have shown a high advantage of using many individual predictors integrated into the ensemble which are responsible for the final forecast. Especially efficient is the application of non-standard wavelet application in the formation of an ensemble, as well as the use of LSTM as the basic prediction unit. The novelty of the paper is the critical comparative analysis of the time series prediction methods applied for load forecasting in the power system. The presented approach may be useful for the users involved in power system operation management.