ABSTRACT A new upper limit on the 21 cm signal power spectrum at a redshift of z ≈ 9.1 is presented, based on 141 h of data obtained with the Low-Frequency Array (LOFAR). The analysis includes significant improvements in spectrally smooth gain-calibration, Gaussian Process Regression (GPR) foreground mitigation and optimally weighted power spectrum inference. Previously seen ‘excess power’ due to spectral structure in the gain solutions has markedly reduced but some excess power still remains with a spectral correlation distinct from thermal noise. This excess has a spectral coherence scale of 0.25–0.45 MHz and is partially correlated between nights, especially in the foreground wedge region. The correlation is stronger between nights covering similar local sidereal times. A best 2-σ upper limit of $Δ ^2_{21} \lt (73)^2\, \mathrm{mK^2}$ at $k = 0.075\, \mathrm{h\, cMpc^{-1}}$ is found, an improvement by a factor ≈8 in power compared to the previously reported upper limit. The remaining excess power could be due to residual foreground emission from sources or diffuse emission far away from the phase centre, polarization leakage, chromatic calibration errors, ionosphere, or low-level radiofrequency interference. We discuss future improvements to the signal processing chain that can further reduce or even eliminate these causes of excess power.